xtuner-eval_niah

📖 Single-Hop & Multi-Hop NIAH

This page provides specific evaluation methods for single-hop and multi-hop needle-in-a-haystack tasks.

---

Preparation

Environment

• Please use a environment with transformers >= 4.45.1 (to use YARN). see niah_requirements.txt.

• And to avoid confict, please don't install XTuner in your environment, I have copy all useful files in xtuner.

Model Parameters

• Please place the pretrained parameters of the model to be tested into ./xtuner/vision_niah/model_weights/.

• Please pay extra attention to modifying the "rope_scaling": null in VideoChat-Flash-Qwen2-7B_res448/config.json to:

"rope_scaling": {
    "type": "yarn",
    "factor": 4.0,
    "original_max_position_embeddings": 32768
  }

Eval Datasets

• Please download the image and videodata for testing from the Huggingface link.
• We place the annotaions of NIAH in vision_niah/data and vision_niah/data_multi.

Haystack Video

• Please place the haystack video of the model to be tested into ./xtuner/vision_niah/data/haystack_videos. The video_haystack.mkv file used in our testing can be accessed via this link.

---

Eval

• When running the needle-in-a-haystack test, please set the xtuner_eval_niah folder as the root directory of the workspace.

single-hop niah

• Please run the following command.

bash vision_niah/flash_eval_xtuner_single.sh

multi-hop niah

• Please run the following command.

bash vision_niah/flash_eval_xtuner_multi.sh