---
base_model: MCG-NJU/VideoChat3-4B
library_name: peft
tags:
  - peft
  - lora
  - video-temporal-grounding
  - mars2-2026
---

# caT-VTG RLVR adapter, step 900

This is the exact complete checkpoint used by Team caT's final MARS2 2026 formal
inference run `videochat3-whisper-inference-20260731-180133-9b3e95`.

- Base model: `MCG-NJU/VideoChat3-4B`
- Training: direct-base RLVR, eight GPUs
- Checkpoint: `checkpoint-step-000900`
- Adapter: LoRA rank 16, alpha 32, dropout 0.0
- Adapter weights SHA-256:
  `fac5e9b635067e3046a6b26362cd5cbf3fc07bc558daaca86964e35b5f917dbd`
- Complete checkpoint content SHA-256:
  `649abf427f52a0859f2eaae4ced5f34374384e52b4dfbb783f921ffc17b33ed1`

The exact checkpoint is stored under `checkpoint-step-000900/` so its original
`README.md` can remain part of the immutable content identity while this repository
keeps a separate Hugging Face model card. `adapter_config.json` and
`adapter_model.safetensors` in that directory are sufficient for the portable
reproduction runner in the companion GitHub repository. The optimizer, rank RNG,
training-state, completion marker, and checkpoint manifest files are included so
reviewers can also verify the original complete checkpoint identity.

The PEFT config is preserved byte-for-byte and therefore contains the absolute base
model path from the training host. The companion loader first materializes the base
model from `MCG-NJU/VideoChat3-4B` and then loads this adapter; it does not depend on
that historical host path. Point the reproduction command at
`model/checkpoint-step-000900` after downloading this repository.

No official leaderboard score is claimed here. Use of the adapter is subject to the
base model's license and the MARS2 competition data terms.