physics-r1-seed17-v4-step40-fsdp

Physics-R1 — Seed 17, v4 step-40 (FSDP-sharded)

Project Page | Paper | Code | Training corpus

Step-40 intermediate checkpoint from the seed-17 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 (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.

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
• Seed-17 row: physics-r1-seed17-canonical-step63-fsdp
• Seed-23 row: physics-r1-seed23-canonical-step60-fsdp

Variants

Training recipe

• Base model: 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 — 2,268 audited closed-form problems
• Hardware: 4×H200 (FSDP1 4-way sharded)

Full hyperparameters in the paper appendix.

• Seed / step: 17 / 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:

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-seed17-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-seed17-v4-step40-fsdp-hf

Then load with standard HF:

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "./physics-r1-seed17-v4-step40-fsdp-hf",
    torch_dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("./physics-r1-seed17-v4-step40-fsdp-hf")

License

Apache 2.0, inheriting from the base model 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

@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}
}