{"id":"arxiv:2512.13961","title":"Olmo 3","date":"2025-12-15","section":"LLMs: architectures, context, training","quality":0.225076,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/","source_url":"https://arxiv.org/abs/2512.13961","text_available":true,"authors":["Olmo Team","Allyson Ettinger","Amanda Bertsch","Bailey Kuehl"],"outgoing":[{"source":"arxiv:2512.13961","target":"arxiv:1511.05952","evidence":"title","id":"arxiv:1511.05952","title":"Prioritized Experience Replay","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c88bf0c11c1817dcc457c001dfef5cd120a0946c8a009a0fe40648dae03eb7d4/"},{"source":"arxiv:2512.13961","target":"arxiv:2304.15004","evidence":"title","id":"arxiv:2304.15004","title":"Are Emergent Abilities of Large Language Models a Mirage?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d7cd8c081ad54f2fdaa2211a4b49dc21f6ff2db7282e287546d210ef90e84323/"},{"source":"arxiv:2512.13961","target":"arxiv:2402.03300","evidence":"id","id":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/636fb871904bc06f742cfdea069796f9e9665f95bd58dc63260eaff5154ef985/"},{"source":"arxiv:2512.13961","target":"arxiv:2405.16684","evidence":"id","id":"arxiv:2405.16684","title":"gzip Predicts Data-dependent Scaling Laws","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8d5456f36b61f2ce9a45da0889b5f0318dc816cab90ee57d54bba06246041de3/"},{"source":"arxiv:2512.13961","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2512.13961","target":"arxiv:2501.00656","evidence":"id","id":"arxiv:2501.00656","title":"2 OLMo 2 Furious","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1ca88dba5d5c59e8cb23b2a304e74370b8ec69077b714a30d426300f7ef2a0b9/"},{"source":"arxiv:2512.13961","target":"arxiv:2501.19393","evidence":"id","id":"arxiv:2501.19393","title":"s1: Simple test-time scaling","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2a88818f7605b71cd0160831faebd51b3274c24ffd29626b4d519390b4bffffe/"},{"source":"arxiv:2512.13961","target":"arxiv:2503.14476","evidence":"id","id":"arxiv:2503.14476","title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/548e934b912ef84691ace63bf1cfd670e7b0257d9d6986dde926f569e7b2f8b7/"},{"source":"arxiv:2512.13961","target":"arxiv:2503.20783","evidence":"title","id":"arxiv:2503.20783","title":"Understanding R1-Zero-Like Training: A Critical Perspective","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b49a056f3ce8e269dd65e3b11605927f3c4d0ef074bc68e0b1d7f84e6c9c6c6a/"},{"source":"arxiv:2512.13961","target":"arxiv:2504.13837","evidence":"id","id":"arxiv:2504.13837","title":"Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b90f70269aeb02311d74ffaadf47e139683424a47118dc18619000ba72174d17/"},{"source":"arxiv:2512.13961","target":"arxiv:2506.10947","evidence":"id","id":"arxiv:2506.10947","title":"Spurious Rewards: Rethinking Training Signals in RLVR","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f48e13a9656d8eff1882eb80988e07a820c4799401298947deac735c004e8592/"},{"source":"arxiv:2512.13961","target":"arxiv:2508.07629","evidence":"id","id":"arxiv:2508.07629","title":"Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/"},{"source":"arxiv:2512.13961","target":"arxiv:2511.07317","evidence":"title","id":"arxiv:2511.07317","title":"RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c0c18cb0ae7fe76a7239f38ce819c7ac588c1184dd27c5672f868ff4e2e26605/"},{"source":"arxiv:2512.13961","target":"url:ai.meta.com/research/publications/code-llama-open-foundation-models-for-code","evidence":"id","id":"url:ai.meta.com/research/publications/code-llama-open-foundation-models-for-code","title":"Code Llama: Open Foundation Models for Code","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/0dc2d6030e514e45150f64f511c0f50fc010fdc1ddae0bdca7edcd189b084d41/"},{"source":"arxiv:2512.13961","target":"url:rlhfbook.com","evidence":"url","id":"url:rlhfbook.com","title":"Reinforcement Learning from Human Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/20e354559fad09ad4d6338d104c9c3f822a48e7fbf86cd8afa1ddf951b223658/"}],"incoming":[{"source":"arxiv:2601.19897","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2601.19897","title":"Self-Distillation Enables Continual Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/"},{"source":"arxiv:2601.20802","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2601.20802","title":"Reinforcement Learning via Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6073ec3d17b9a932a68e79ae078928f4635e40fc79cfcbc31709715d1f1e1ffa/"},{"source":"arxiv:2602.18037","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2602.18037","title":"Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/75161700a0f3e03055116a289c82e9f9fa7d8fee7a6f802e116216eac3a70e31/"},{"source":"arxiv:2603.24472","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2603.24472","title":"Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/"},{"source":"arxiv:2604.02288","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2604.02288","title":"Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/"},{"source":"arxiv:2606.03982","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2606.03982","title":"Language Models Compare Quantities Using Number-specific and Unit-specific Heuristics","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48f07d03c9308ca8fb43ecd2ac22978f6dcbd4bc081056f022ca511711942dbf/"},{"source":"arxiv:2608.13040","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2608.13040","title":"Latent On-Policy Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/"},{"source":"arxiv:2610.02140","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2610.02140","title":"Finetuning with Sampling: SFT Learns Better Than You Think","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/"},{"source":"url:cameronrwolfe.substack.com/p/grpo-tricks","target":"arxiv:2512.13961","evidence":"id","id":"url:cameronrwolfe.substack.com/p/grpo-tricks","title":"GRPO++: Tricks for Making RL Actually Work (Cameron Wolfe)","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d170bc3da5b8e6a055fd511719fcc3764635236bac638947bd0a018412245eb5/"}],"text_parts":[{"start":0,"end":15710,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html"},{"start":15710,"end":29482,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-2.html"},{"start":29482,"end":45474,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html"},{"start":45474,"end":61394,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html"},{"start":61394,"end":77384,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html"},{"start":77384,"end":93008,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html"},{"start":93008,"end":108755,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html"},{"start":108755,"end":124702,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html"},{"start":124702,"end":140386,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html"},{"start":140386,"end":155608,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html"},{"start":155608,"end":171544,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html"},{"start":171544,"end":186919,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html"},{"start":186919,"end":202849,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html"},{"start":202849,"end":218665,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html"},{"start":218665,"end":234468,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html"},{"start":234468,"end":250350,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html"},{"start":250350,"end":265504,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html"},{"start":265504,"end":279647,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html"},{"start":279647,"end":288421,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-19.html"}],"contents":[{"title":"Olmo 3","offset":3752,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"Abstract","offset":3762,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"1 Introduction","offset":4316,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"2 Model Flow for Olmo 3","offset":8351,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"2.1 Base Model Training","offset":8879,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"Evaluation","offset":9177,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"Data curriculum","offset":9976,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"Open artifacts","offset":12271,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"2.2 Post-training","offset":12997,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-1.html#text"},{"title":"2.3 Results","offset":18116,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-2.html#text"},{"title":"2.4 Costs","offset":19244,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-2.html#text"},{"title":"3 Olmo 3 Base","offset":22703,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-2.html#text"},{"title":"3.1 Main Results for Olmo 3 Base","offset":24411,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-2.html#text"},{"title":"3.2 Modeling and Architecture","offset":32952,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"Architecture","offset":33142,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"Training","offset":33834,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"Tokenizer","offset":35123,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.3 Experimental Design and Evaluation","offset":35283,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.3.1 Clustering Tasks","offset":37484,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.3.2 Scaling analysis","offset":38798,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.3.3 Signal-to-Noise Analysis","offset":40341,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.3.4 OlmoBaseEval","offset":41655,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.4 Stage 1: Pretraining","offset":43018,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"3.4.1 Preparing our Web Data Pool","offset":45249,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"Text extraction","offset":45451,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-3.html#text"},{"title":"Heuristic filtering","offset":45977,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Deduplication","offset":47181,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Topic and quality classification","offset":50665,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Final web data pool","offset":52105,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"3.4.2 Preparing our olmOCR science PDFs Data Pool","offset":52550,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"olmOCR text extraction","offset":53170,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Deduplication","offset":53791,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"PII filtering","offset":54280,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Heuristic filtering","offset":56075,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"3.4.3 Preparing Code, Math, and other sources","offset":56766,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Code","offset":56818,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Math","offset":57124,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Other","offset":57938,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"3.4.4 Sampling and Mixing over Data Pools","offset":58427,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Constrained data mixing","offset":59357,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-4.html#text"},{"title":"Quality-aware upsampling","offset":64695,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"Evaluation during pretraining","offset":67248,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"3.5 Stage 2: Midtraining","offset":67876,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"3.5.1 Methodological framework","offset":69295,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"Targeted capability boosts","offset":69332,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"Integration tests","offset":71636,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"3.5.2 Capability Improvements for Final Data Mix","offset":72955,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"Math capabilities","offset":73915,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-5.html#text"},{"title":"Code capabilities","offset":77384,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"QA and knowledge access capabilities","offset":79671,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"Cross-Capability instruction data","offset":81822,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"Cross-capability thinking traces","offset":82873,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"High quality web and PDF data","offset":87638,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"3.5.3 Decontamination","offset":89480,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"Method and tooling","offset":90158,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"3.5.4 Key findings","offset":92191,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"Candidate mix quality improves over time","offset":92477,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-6.html#text"},{"title":"Performance shows substantial domain tradeoffs","offset":93422,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Thinking/instruct data benefits base performance","offset":96070,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Leave special tokens for SFT stage","offset":96838,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Extent and impact of decontamination are variable","offset":98394,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Model souping can improve midtraining performance","offset":100883,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"3.6 Stage 3: Long-context Extension","offset":101825,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Long-context extension strategy","offset":102538,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"High variance in open-model recipes","offset":102939,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Olmo 3 long-context recipe","offset":104290,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Overall results","offset":105307,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"3.6.1 Sourcing Long Context Data","offset":106328,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"olmOCR science PDFs","offset":106367,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Data filtering","offset":106671,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"3.6.2 Experiments with Synthetic Augmentation","offset":108186,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Generation pipeline","offset":108728,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-7.html#text"},{"title":"Synthetic aggregation tasks","offset":110061,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"3.6.3 Choosing Data Mix and Token Budget","offset":111137,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Interleaving long- and short-context data","offset":111184,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Longer extension helps","offset":111693,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"3.6.4 Curating a Training Recipe for Extension","offset":112039,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"RoPE extension","offset":112092,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Document packing","offset":112698,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Intra-document masking","offset":113348,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"LC training infrastructure","offset":113768,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Model souping","offset":114281,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"3.7 Base Model Results","offset":114697,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4 Olmo 3 Think","offset":115257,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4.1 Main Results for Olmo 3 Think","offset":118026,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4.1.1 Evaluation Details","offset":118065,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4.1.2 Main Results","offset":122195,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4.2 Supervised Finetuning with Dolci Think SFT","offset":124224,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"4.2.1 Dolci Think SFT : Data Curation","offset":124658,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-8.html#text"},{"title":"Step 1: sourcing prompts and generating reasoning traces","offset":125062,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 2: filtering","offset":127906,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 3: data mixing","offset":128994,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 4: decontamination","offset":129952,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.2.2 Training","offset":130726,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.3 Preference Tuning with Delta Learning","offset":131522,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Delta Learning","offset":133290,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.3.1 Dolci Think -DPO: Preference Data Creation","offset":133891,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 1: sourcing prompts and contrastive completions","offset":135019,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 2: filtering","offset":135544,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"Step 3: mixing","offset":135902,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.3.2 Training","offset":136736,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.4 Reinforcement Learning with OlmoRL : The Cherry on Top","offset":137387,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"4.4.1 OlmoRL Algorithmic Details","offset":138265,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-9.html#text"},{"title":"OlmoRL formulation","offset":140741,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Verifiers","offset":143060,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"4.4.2 Dolci-Think-RL : Curating a State-of-the-art RLVR Dataset","offset":145523,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Step 1: sourcing prompts","offset":146229,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Step 2: offline difficulty filtering","offset":149782,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Step 3: data mixing","offset":150645,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"4.4.3 OlmoRL Infrastructure in Open Instruct","offset":152439,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Fully asynchronous training","offset":153845,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Continuous batching","offset":155581,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-10.html#text"},{"title":"Active sampling","offset":156544,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Inflight updates","offset":157335,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Better threading and engineering","offset":158155,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"4.5 Key Findings","offset":158992,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"DPO yields gains where SFT on the same data cannot","offset":159014,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"DPO and SFT both benefit from RL, but DPO remains a better starting point","offset":160019,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Rewards steadily increase across all domains during RL","offset":161544,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Mixing RL data from varied domains can prevent over-optimization","offset":162183,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Mixing data yields lower train reward, but not lower downstream performance","offset":162870,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Continuous batching and inflight updates are crucial to training speed","offset":163578,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"OlmoRL shows significant improvement in precise instruction following","offset":165158,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"5 Olmo 3 Instruct","offset":165539,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"5.1 Main Results for Olmo 3 Instruct","offset":166497,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"5.2 Supervised Finetuning with Dolci Instruct SFT","offset":170861,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"5.2.1 Function-calling Training Data","offset":171098,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Trajectories with real interactions","offset":171501,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-11.html#text"},{"title":"Trajectories with simulated interactions","offset":173368,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Balancing function diversity with interaction complexity","offset":174547,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Unified data format","offset":175211,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Evaluating function calling","offset":176079,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.2.2 Curating Dolci Instruct SFT","offset":178522,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Step 1. Sourcing Prompts and Completions","offset":178562,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Step 2: Filtering & Mixing","offset":179170,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Starting from Olmo 3 Think SFT","offset":179468,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.3 Preference Tuning with Dolci Instruct DPO","offset":179757,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.3.1 Preference Signals","offset":180525,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Delta-learning heuristic pairs","offset":180691,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Delta-aware GPT-judged pairs","offset":181029,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Multi-turn preferences","offset":182179,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"Controlling length bias","offset":183192,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.3.2 Prompt Mixing","offset":184314,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.3.3 Training","offset":186060,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.4 Reinforcement Learning with Dolci Instruct -RL","offset":186524,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.4.1 Training","offset":186898,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-12.html#text"},{"title":"5.5 Key Findings","offset":188430,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Starting from the Olmo 3 Think SFT is helpful","offset":188539,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"High contrast in preference pairs drives DPO improvements","offset":188910,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Combining different preference signals improves overall performance","offset":189969,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"The ideal amount of preference data depends on the downstream task","offset":190388,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Concise, usable model outputs from preference tuning can boost RL performance","offset":191208,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Need for tools","offset":192726,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"6 Olmo 3 RL-Zero","offset":193596,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"6.1 Reinforcement Learning From Base with Dolci RL-Zero","offset":195373,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Data","offset":195434,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Prompt and eval template","offset":196368,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"RL algorithm","offset":197009,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"6.2 Key Findings","offset":197377,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Olmo 3 RL-Zero can strongly improve on reasoning","offset":197399,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Olmo 3 RL-Zero mix can benchmark challenges in multi-objective RL","offset":198333,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Olmo 3 RL-Zero can benchmark reasoning data mixes in midtraining","offset":198931,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Active sampling stabilizes training","offset":199657,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Eval decontamination is verified via spurious rewards","offset":200218,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Appendix A Appendix","offset":201724,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Author Contributions","offset":201748,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-13.html#text"},{"title":"Acknowledgments","offset":204675,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.1 Base Model Additional Training Details","offset":205213,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2 Base Model Additional Data Details: Pretraining","offset":205751,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2.1 CommonCrawl","offset":205808,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2.2 Deduplication","offset":208487,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Exact Deduplication","offset":209438,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"MinHash Fuzzy deduplication","offset":210276,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Suffix Array deduplication","offset":212060,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2.3 Topic Classification","offset":213260,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2.4 CommonCrawl Mixing","offset":213971,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Problem formulation","offset":214631,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Parameterizing the solution space","offset":215223,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Implementation details","offset":217120,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"A.2.5 Validating Quality Upsampling and Mixing","offset":217640,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Targeted mixing","offset":217897,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Quality-aware upsampling","offset":218633,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-14.html#text"},{"title":"Reconciling upsampling and mixing","offset":219328,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.3 Base Model Additional Data Details: Midtraining","offset":221152,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.3.1 Math Capabilities","offset":221425,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"TinyMATH","offset":222152,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"CraneMath","offset":223136,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"MegaMatt","offset":224658,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"OMR Rewrites","offset":225671,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"Key Findings and Results","offset":225911,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.3.2 Code Capabilities","offset":227744,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"CraneCode","offset":228371,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.3.3 Thinking Capabilities","offset":229248,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"Meta-reasoning","offset":229282,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"Existing thinking traces","offset":231807,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.4 Base Model Additional Evaluation Details","offset":233322,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"Expanding OLMES tasks","offset":233688,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-15.html#text"},{"title":"A.4.1 Base Evaluation suites","offset":235392,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Base Easy suite","offset":236179,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Base Main suite","offset":237220,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Base Chat suite","offset":238234,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Base Long-Context suite","offset":238937,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Base Held-out Suite","offset":239737,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.4.2 New Evaluation Benchmarks","offset":240067,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Basic Skills","offset":240105,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Gen2MC","offset":240652,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Masked perplexity","offset":241696,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.5 Base Model Additional Decontamination Details","offset":245348,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.5.1 Definitions and Preliminaries","offset":245786,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.5.2 Example of Contamination","offset":246604,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.5.3 Eval Normalization","offset":247307,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"A.5.4 Decon Implementation","offset":248481,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Detecting contamination","offset":248733,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Cluster discovery","offset":248927,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"IDF-weighted overlap","offset":250113,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-16.html#text"},{"title":"Cluster match length decay","offset":250500,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Cluster discovery threshold","offset":251074,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Non-comparative","offset":251381,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Inverted index","offset":251833,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Hot n-grams","offset":252524,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.5.5 Scoring System","offset":253077,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Length penalty","offset":253408,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Confidence adjusted weight","offset":254012,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Base scores","offset":254977,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Answer proximity","offset":255574,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Passage proximity","offset":255777,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.6 Post-Training Additional Training Details","offset":256063,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.6.1 Supervised Finetuning Details","offset":256114,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Using OLMo-core infrastructure for SFT Training","offset":256156,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.6.2 Preference Tuning Details","offset":257268,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Training Settings","offset":257306,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.6.3 Reinforcement Learning Details","offset":259050,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.6.4 RL-Zero Details","offset":259788,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.7 Post-Training Additional Data Details","offset":261737,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.7.1 Filtering for Dolci Think -SFT","offset":261784,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"A.7.2 Tool-use data","offset":264680,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Additional details about the Science QA dataset","offset":264706,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Additional details about the Web Search QA dataset","offset":265446,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-17.html#text"},{"title":"Additional details about simulated interaction trajectories","offset":266614,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"A.7.3 Coding Data Synthesis Pipeline","offset":267198,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"A.7.4 Dolci Instruct DPO Details","offset":268488,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"DPO prompt mixing","offset":268527,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Model pool for LLM-judged pairs","offset":268604,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"A.8 Post-Training Additional Evaluation Details","offset":270470,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"A.8.1 General Evaluation Settings","offset":270523,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Is AlpacaEval useful?","offset":271966,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"A.8.2 Safety Evaluations Overview","offset":275539,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Development safety evaluations","offset":276084,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Unseen safety evaluations","offset":276387,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Averaging and reported metrics","offset":276677,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-18.html#text"},{"title":"Instructions for reporting errors","offset":287457,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/part-19.html#text"}],"text_code_points":288421,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/source.md"}