RSI-Qwen

RSI-Qwen

Autonomous research loop for improving Qwen/Qwen3-VL-4B-Instruct on the official HuggingFaceM4/DocumentVQA benchmark.

The repo is designed for agentic training research: the benchmark and evaluator stay fixed, while an agent iterates on train.py, runs training, measures the result on the full validation split, and keeps only real gains. The project is inspired by karpathy/autoresearch, but scoped to a concrete public VLM benchmark with a reproducible contract.

If this project is useful for your research, evals, or agent workflows, please star the repo.

Branches

This repository now contains the two previously separate codebases as different branches:

Branch	Target hardware	Status
main	NVIDIA / CUDA multi-GPU	Primary branch. Uses torchrun, supports DeepSpeed configs, and is the recommended branch for fast experiment cycles.
mlx	Apple Silicon / MPS	Historical branch imported from the former autoresearch-qwen-mlx repository and preserved here as the Mac-focused variant.

Use the README on each branch for branch-specific commands. On main the entrypoint is ./run_experiment.sh; on mlx it is uv run python run_experiment.py.

Why This Repo Exists

• Fixed benchmark: full official DocVQA train, validation, and test splits
• Fixed evaluator: validation score is always computed by the repository evaluator
• One mutable surface: agents are expected to edit train.py
• Reproducible loop: prepare, train, evaluate, keep or discard, repeat
• Public benchmark mindset: improvements should come from better training decisions, not from moving the goalposts

Benchmark Contract

Component	Contract
Base model	Qwen/Qwen3-VL-4B-Instruct
Dataset	HuggingFaceM4/DocumentVQA official splits
Training split	Full train split
Validation split	Full validation split
Test split	Full blind test split
Metric	Mean ANLS on the full validation split
Mutable file	train.py
Fixed files	evaluate.py, src/, benchmark contract, submission tooling

More benchmark details are documented in benchmarks/README.md.

How The Loop Works

prepare.py          Download dataset + model snapshot
      |
train.py            Mutable training code (the agent edits this)
      |
evaluate.py         Fixed validation evaluator / blind test exporter
      |
run_experiment.sh   One full train -> eval iteration on main

The only objective is to maximize val_score, defined as mean ANLS on the full official validation split.

Quick Start (main)

curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
uv run python prepare.py
uv run autoresearch-qwen doctor
uv run python evaluate.py --base-only --split validation
./run_experiment.sh | tee run.log

Useful follow-up commands:

uv run python analysis.py
uv run python submit_test.py

Repository Layout

train.py                            Mutable training code
evaluate.py                         Fixed evaluator
run_experiment.sh                   One-command train -> eval pipeline on main
analysis.py                         Result visualization
prepare.py                          Dataset + model downloader
submit_test.py                      Blind test export + submission packaging
check_submission.py                 Submission validator
program.md                          Full agent protocol
benchmarks/README.md                Benchmark definition
configs/                            DeepSpeed configs for multi-GPU runs
src/autoresearch_qwen/              Fixed library code

Running An Agent

The full experiment protocol lives in program.md. A practical starting prompt is:

Read the entire repository, especially README.md and program.md. You may read all files for context, but only edit train.py. Run `uv run autoresearch-qwen doctor --json`, record a `--base-only` validation baseline, then start the autoresearch loop. Parse `artifacts/last_result.json` after each run and keep only changes that improve val_score.

Results, Analysis, and Submission

• artifacts/last_result.json stores the latest train/eval result payload
• analysis.py plots experiment progress from accumulated results
• submit_test.py exports predictions for the blind DocVQA test split
• check_submission.py validates a submission bundle locally before upload

Contributing

Issues and pull requests are welcome, especially for:

• stronger training recipes that respect the benchmark contract
• better experiment tooling and reproducibility
• clearer docs and onboarding
• hardware-specific improvements that belong on a dedicated branch

If you want to change the benchmark contract itself, open an issue first so the rationale is explicit.

Acknowledgements

• karpathy/autoresearch for the original autonomous research-loop framing
• Qwen for the base vision-language model
• Hugging Face M4 for the public DocVQA dataset release

Star Trend

License

MIT