{"id":"arxiv:2608.27370","title":"Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090","date":"2026-08-27","section":"Data, training, optimization","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/","source_url":"https://arxiv.org/abs/2608.27370","text_available":true,"authors":["Kairong Luo","Jiarui Cui","Yaorui Yin","Shengqi Chen"],"outgoing":[],"incoming":[],"text_parts":[{"start":0,"end":15588,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html"},{"start":15588,"end":31304,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html"},{"start":31304,"end":46994,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html"},{"start":46994,"end":62401,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html"},{"start":62401,"end":78036,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html"},{"start":78036,"end":93560,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html"},{"start":93560,"end":98539,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-7.html"}],"contents":[{"title":"Puro-2B : Poor Lab’s Qwen2-1.5B Trained on RTX 5090 within $5090 Thanks: Corresponding authors.","offset":83,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"摘要","offset":182,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"1 Introduction","offset":2266,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"2 Overview","offset":11876,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"2.1 Pipeline at a Glance","offset":12591,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"Model architecture and pretraining system ( Sections 3.1 and 3.2 ).","offset":12621,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"Two-phase pretraining, curriculum, and optimization ( Sections 3.3 and 3.4 ).","offset":13852,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-1.html#text"},{"title":"Data acquisition, filtering, and mixture ( Section 3.6 ).","offset":16598,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Evaluation and post-training ( Sections 3.5 and 4 ).","offset":18691,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"2.2 Reproduction-Cost Boundary","offset":19692,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"3 Training Recipe","offset":21201,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"3.1 Training Infrastructure","offset":21607,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"3.1.1 RTX 5090 as a Cost-Effective GPU Choice","offset":21640,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"3.1.2 Hardware Setup and Tweaks","offset":25539,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Intra-node bandwidth.","offset":25955,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Inter-node bandwidth.","offset":27672,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Caveats.","offset":29279,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"3.1.3 Efficient Training System","offset":29885,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Communication-aware parallelism.","offset":30403,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Appropriate micro-batch size.","offset":31267,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-2.html#text"},{"title":"Overcoming load imbalance.","offset":32321,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"3.2 FP8 Mixed-Precision Training","offset":35418,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"3.3 Hyperball Optimization","offset":38612,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"3.3.1 Effective Learning Rate","offset":40127,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"Scale invariance motivates the effective learning rate.","offset":40560,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"MuonH makes the effective LR schedule explicit.","offset":41939,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"Matching the effective LR schedule also aligns validation loss.","offset":43386,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"Effective LR provides a more informative descriptor of the loss trend.","offset":45765,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"Effective LR MPL helps interpret the late-stage crossover.","offset":46928,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-3.html#text"},{"title":"3.3.2 Learning Rate Schedule Design","offset":48306,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Continual training motivates an open-ended schedule followed by a controlled terminal decay.","offset":48348,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Observation 1: a larger peak LR calls for a longer decay.","offset":49383,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Observation 2: a longer training horizon also calls for a longer decay.","offset":50382,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"The production schedule combines these three practical features.","offset":51082,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Multi-power law provides a back-of-the-envelope, cross-schedule estimator under limited compute.","offset":51900,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"3.4 Curriculum Model Averaging","offset":52625,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"The data curriculum arranges the order within each component while keeping the target mixture weight.","offset":54464,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"The canonical CMA run uses the best-performing late-stage continuation with model averaging.","offset":56109,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Ablation study on curriculum, model average, and const-LR continuation.","offset":57299,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"3.5 Post-Training Recipe","offset":58878,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Motivation and question.","offset":58908,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Data construction and experimental setup.","offset":60614,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"Evaluation.","offset":62382,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-4.html#text"},{"title":"3.6 Data Recipe","offset":63216,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"3.6.1 How Do We Select a Good Data Recipe?","offset":64440,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Fine-grained data slice selection.","offset":64489,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Feature-guided data selection.","offset":65554,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Capability balance across domains.","offset":66536,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"3.6.2 How Do We Preprocess Data and Reproduce the Shards?","offset":67767,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Data preprocessing pipeline.","offset":67831,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Kaiyuan-Spark preprocessing framework.","offset":69061,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Two-phase data and transition schedule.","offset":70153,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4 Evaluation","offset":71303,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.1 Evaluation Setup","offset":71320,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.1.1 Model Selection","offset":71346,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.1.2 Benchmarks","offset":74457,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.1.3 Implementation Details","offset":75585,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.2 Resulting Performance","offset":77233,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Mathematics and code.","offset":77264,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"Reasoning and knowledge.","offset":78004,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-5.html#text"},{"title":"4.3 Cost Estimation","offset":78825,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"4.3.1 Cost-Saving Factors","offset":78850,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"RTX 5090 exhibits a high cost-performance ratio and utilization.","offset":80612,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"MuonH provides a quality-equivalent compute shift in the matched ladder.","offset":81064,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"Blockwise FP8 retains a stable quality fraction while accelerating the 2B proxy.","offset":81804,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"The Puro Cost Scaling Law summarizes recipe-specific scale-down behavior.","offset":82739,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"4.3.2 Cross-Model Reproduction Cost","offset":84715,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"Estimation protocol.","offset":84757,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"Pareto-frontier result.","offset":85830,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"4.4 Post-Training Results and Analysis","offset":86135,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"GSM8K-based SFT.","offset":86830,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"Math&Code SFT with replay.","offset":87564,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"Tulu-3 mixed-domain SFT.","offset":88420,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"5 Related Works","offset":89875,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"5.1 Open-Recipe Language Models","offset":89895,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"5.2 Low-Cost Language Model Pretraining","offset":93073,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-6.html#text"},{"title":"6 Conclusion and Future Direction","offset":96141,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-7.html#text"},{"title":"7 Acknowledgments","offset":97934,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/part-7.html#text"}],"text_code_points":98539,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e7c6bd97e97892c4e1efe65bec9c8dd703a7e817d9544e8b27822583311ab5ab/source.md"}