--- license: apache-2.0 language: - en - zh base_model: Qwen/Qwen3-VL-8B-Thinking tags: - physics - reasoning - multimodal - rl - grpo - verl - fsdp - arxiv:2605.14040 pipeline_tag: image-text-to-text library_name: transformers --- # Physics-R1 — Seed 42, v4 step-40 (FSDP-sharded) [**Project Page**](https://shanyang.me/physics-r1-page/) | [**Paper**](https://huggingface.co/papers/2605.14040) | [**Code**](https://github.com/shanyang-me/physics-r1-neurips2026) | [**Training corpus**](https://huggingface.co/datasets/shanyangmie/physr1corp) **Step-40 intermediate checkpoint** from the seed-42 v4 (audited-data) Physics-R1 training run. Released for training-curve ablations and reproducibility. Fine-tune of `Qwen3-VL-8B-Thinking` on the audited [`PhysR1Corp`](https://huggingface.co/datasets/shanyangmie/physr1corp) (2,268 closed-form physics problems) via full-parameter FSDP1 GRPO with binary correctness reward. Released alongside [Physics-R1: An Audited Olympiad Corpus and Recipe for Visual Physics Reasoning](https://huggingface.co/papers/2605.14040). ## Performance Intermediate-step checkpoint released for training-curve ablations. **Not the paper-headline checkpoint** — for exact Table 2 numbers see: - Seed-42 row: [`physics-r1-seed42-v4-step60-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed42-v4-step60-fsdp) - Seed-17 row: [`physics-r1-seed17-canonical-step63-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed17-canonical-step63-fsdp) - Seed-23 row: [`physics-r1-seed23-canonical-step60-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed23-canonical-step60-fsdp) ## Variants | Checkpoint | Used for | Notes | |---|---|---| | [`physics-r1-seed42-v4-step60-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed42-v4-step60-fsdp) | Paper Table 2 seed-42 row | Binary, step 60 | | [`physics-r1-seed17-canonical-step63-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed17-canonical-step63-fsdp) | Paper Table 2 seed-17 row | Binary, step 63 | | [`physics-r1-seed23-canonical-step60-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed23-canonical-step60-fsdp) | Paper Table 2 seed-23 row | Binary, step 60 | | [`physics-r1-seed17-v4-step60-fsdp`](https://huggingface.co/shanyangmie/physics-r1-seed17-v4-step60-fsdp) | Seed-17 v4 re-validation, tracks canonical | Binary, step 60 | | [`physics-r1-seed42-v4-step{40,50}-fsdp`](https://huggingface.co/shanyangmie) | Step ablation (seed 42) | Earlier steps | | [`physics-r1-seed17-v4-step{40,50}-fsdp`](https://huggingface.co/shanyangmie) | Step ablation (seed 17) | Earlier steps | ## Training recipe - **Base model**: [`Qwen/Qwen3-VL-8B-Thinking`](https://huggingface.co/Qwen/Qwen3-VL-8B-Thinking) - **Algorithm**: GRPO (verl 0.6.1, full-parameter FSDP1 — `actor.strategy=fsdp`, *not* `fsdp2`; FSDP2 fails on Qwen3-VL visual encoder device placement) - **Reward**: binary correctness, per-subpart Sonnet judge with problem-level AND aggregation (see paper §3.2) - **Data**: [`shanyangmie/physr1corp`](https://huggingface.co/datasets/shanyangmie/physr1corp) — 2,268 audited closed-form problems - **Hardware**: 4×H200 (FSDP1 4-way sharded) Full hyperparameters in the paper appendix. - **Seed / step**: 42 / 40 ## Format: verl FSDP-sharded checkpoint (conversion required) This checkpoint is saved in **verl's FSDP-sharded format**, not safetensors. It is **not** directly loadable via `AutoModelForImageTextToText.from_pretrained` without a merge step. ### File layout ``` actor/ ├── huggingface/ # HF-style config + tokenizer ├── model_world_size_4_rank_{0,1,2,3}.pt # 4-way FSDP weight shards (~8.7 GB each, ~35 GB total) ├── optim_world_size_4_rank_{0,1,2,3}.pt # optimizer state (~17.5 GB each, not needed for inference) ├── extra_state_world_size_4_rank_{0..3}.pt └── fsdp_config.json data.pt # verl bookkeeping (not needed for inference) ``` ### Convert to HF safetensors Use verl's `model_merger.py`: ```bash git clone https://github.com/volcengine/verl cd verl # Download only the inference-required files (skips ~70 GB of optimizer state) huggingface-cli download shanyangmie/physics-r1-seed42-v4-step40-fsdp \\ --include "actor/model_world_size_4_rank_*.pt" \\ --include "actor/huggingface/*" \\ --include "actor/fsdp_config.json" \\ --include "actor/extra_state_world_size_4_rank_*.pt" \\ --local-dir ./ckpt # Merge FSDP shards into HF safetensors python scripts/model_merger.py merge \\ --backend fsdp \\ --hf_model_path Qwen/Qwen3-VL-8B-Thinking \\ --local_dir ./ckpt/actor \\ --target_dir ./physics-r1-seed42-v4-step40-fsdp-hf ``` Then load with standard HF: ```python from transformers import AutoModelForImageTextToText, AutoProcessor model = AutoModelForImageTextToText.from_pretrained( "./physics-r1-seed42-v4-step40-fsdp-hf", torch_dtype="bfloat16", device_map="auto", ) processor = AutoProcessor.from_pretrained("./physics-r1-seed42-v4-step40-fsdp-hf") ``` ## License Apache 2.0, inheriting from the base model [`Qwen3-VL-8B-Thinking`](https://huggingface.co/Qwen/Qwen3-VL-8B-Thinking). Training data (`physr1corp`) is CC BY-NC 4.0, so this derivative checkpoint is intended for **non-commercial research use**. ## Citation ```bibtex @misc{yang2026physicsr1, title = {Physics-R1: An Audited Olympiad Corpus and Recipe for Visual Physics Reasoning}, author = {Yang, Shan}, year = {2026}, url = {https://huggingface.co/papers/2605.14040} } ```