TimeLens2-8B-SFT

This repository contains the supervised fine-tuning (SFT) checkpoint for TimeLens2-8B, a video multimodal language model for temporal grounding. Given a video and a text query, the model is trained to return all time intervals that contain the requested visual evidence.

It is initialized from Qwen3-VL-8B-Instruct and fine-tuned with long-context temporal-grounding supervision. This checkpoint is the starting point for TimeLens2 rollout generation and GRPO training.

For the released TimeLens2 model, see TimeLens2-8B.

Training data and recipe

SFT uses examples from TimeLens2-93K, TimeLens-100K, and the official Ego4D-NLQ-v2 training split. Training retains full-video context and supervises variable-cardinality temporal interval sets. The model sees diverse grounding instructions and timestamp response formats.

The 8B SFT recipe trains for one epoch with packed sequences up to 100K tokens, a global batch size of 256, and a learning rate of 5e-6. The visual encoder is frozen during training.

Intended use

• Generate off-policy rollouts for the TimeLens2 GRPO stage.
• Reproduce or extend the SFT-to-GRPO training pipeline.
• Study long-context video temporal grounding before reinforcement learning.

This is an SFT checkpoint. It is not the final GRPO-optimized release reported in the TimeLens2 paper.

Loading the checkpoint

This checkpoint is hosted in a Hugging Face dataset repository. Download it locally, then load the directory with Transformers:

from huggingface_hub import snapshot_download
from transformers import AutoModelForImageTextToText, AutoProcessor

model_dir = snapshot_download(
    repo_id="MCG-NJU/TimeLens2-8B-SFT",
    repo_type="dataset",
)
model = AutoModelForImageTextToText.from_pretrained(
    model_dir,
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_dir)

See the official repository for the SFT and GRPO training recipes, data annotations, rollout configuration, and evaluation code.

Citation

@article{zhu2026timelens2,
  title={TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs},
  author={Zhu, Yuhan and Ma, Changlian and Zeng, Xiangyu and Li, Xinhao and Zhang, Zhiqiu and Li, Songze and Zhang, Jun and Jiang, Tianxiang and Yang, Yuandong and Yan, Ziang and others},
  journal={arXiv preprint arXiv:2607.17423},
  year={2026}
}