Skip to content

Navigation Menu

Sign in
Sign up

Repository files navigation

VRM-7B

Open-weights 7-8B Visual Reasoning Model. SFT cold-start + GRPO RL post-training on Qwen/Qwen2.5-VL-7B-Instruct.

Documentation

Quickstart (local dev)

git clone https://github.com/tech-sumit/VRM-model.git && cd VRM-model
cp .env.example .env # fill in tokens
make sync # install deps (no GPU needed for dev)
make test # unit tests
make lint # ruff + format check
make typecheck # pyright

Quickstart (RunPod operations)

# 1. Build dataset shards (CPU pod + teacher API)
make data DATA_VERSION=v1
# 2. Stage 1 SFT (~24h on 8xH200)
make train-sft DATA_VERSION=v1 RUN_NAME=sft-2026年05月03日
# 3. Stage 2 GRPO (~10-14d on 8xH200)
make train-grpo \
 SFT_CHECKPOINT=tech-sumit/vrm-7b-sft-2026年05月03日 \
 DATA_VERSION=v1 \
 RUN_NAME=grpo-2026年05月04日
# 4. Eval
make eval CHECKPOINT=tech-sumit/vrm-7b-grpo-2026年05月04日 SUITE=full
# 5. Release (annotated tag triggers vrm-release.yml)
git tag -a vrm-7b-v1.0.0 -m "source=tech-sumit/vrm-7b-grpo-2026年05月04日"
git push origin vrm-7b-v1.0.0

CI/CD via GitHub Actions: see .github/workflows/vrm-*.yml.

License

Apache-2.0.

About

VRM-7B: open-weights visual reasoning model. SFT + GRPO RL on Qwen2.5-VL-7B-Instruct.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

AltStyle によって変換されたページ (->オリジナル) /