You can deploy this model and start earning money today!

WuuuuJH/
ChronusOmni-Purned-7.98b-recover

Name: ChronusOmni-Purned-7.98b-recover
Availability: InStock

ChronusOmni

video-temporal-grounding

Views

No views yet

ChronusOmni 7.984B Pruned + SFT/GRPO Recovered Checkpoint

This repository contains the portable model package for our pruned and recovered ChronusOmni checkpoint. The corresponding code, pruning scripts, SFT recovery scripts, GRPO training scripts, inference scripts, and evaluation utilities are maintained in the GitHub repository:

https://github.com/Q-if/Multi-scale-Omni-Aggregation-Model

What Is Included

The main checkpoint is a native ChronusOmni structural-pruning checkpoint in the following format:

chronusomni-native-mlp-pruning-v2

It stores:

model_state_dict
mlp_intermediate_widths
pruning metadata
SFT/GRPO recovery metadata

It is not a standalone generic Qwen checkpoint and should be loaded with the ChronusOmni pruned-inference script from the GitHub repository.

Expected Repository Layout

Please keep the Hugging Face repository in this layout:

text

.
├── README.md
├── runtime/
│   └── ChronusOmni/
│       ├── config.json
│       ├── BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt
│       ├── large-v3.pt
│       └── ...
├── weights/
│   └── chronusomni_7p984b_teacher256_aux510_iou_grpo_2gpu_s48_ga8_lr5e7_b008.pt
├── wavtemp/
├── packaging/
│   └── chronusomni_portable_manifest.json
├── scripts/
│   └── packaging/
│       ├── patch_chronusomni_relative_paths.py
│       └── validate_chronusomni_package.py
└── docs/
    └── CHRONUSOMNI_PORTABLE_PATHS.md

Large model/runtime files are hosted here. Source code is hosted on GitHub. The small packaging/, scripts/packaging/, and docs/ files are duplicated here so that a downloaded model package can be validated and patched without requiring manual path edits.

Model Checkpoint

Primary checkpoint:

weights/chronusomni_7p984b_teacher256_aux510_iou_grpo_2gpu_s48_ga8_lr5e7_b008.pt

This checkpoint was produced from the 7.984B structurally pruned ChronusOmni model, followed by LoRA/SFT recovery and temporal-IoU GRPO refinement.

Important Path Note

The original runtime/ChronusOmni/config.json may contain machine-specific absolute paths, for example:

json

"music_encoder": "/mnt/shared/public_model/llm/ChronusOmni/BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt",
"speech_encoder": "/mnt/shared/public_model/llm/ChronusOmni/large-v3.pt",
"mm_vision_tower": "/123"

After downloading this repository to a new machine, validate and patch the runtime paths before inference.

bash

PACKAGE_ROOT=/path/to/ChronusOmni-Purned-7.98b-recover

python scripts/packaging/validate_chronusomni_package.py \
  --package_root "$PACKAGE_ROOT" \
  --manifest packaging/chronusomni_portable_manifest.json \
  --allow_missing_vision_tower

python scripts/packaging/patch_chronusomni_relative_paths.py \
  --package_root "$PACKAGE_ROOT" \
  --manifest packaging/chronusomni_portable_manifest.json \
  --allow_missing_vision_tower

--allow_missing_vision_tower is used because some ChronusOmni releases keep the vision-tower path as an internal placeholder. If your local package contains a real local vision-tower directory, you may remove this flag.

Inference

Clone the code repository first:

bash

git clone https://github.com/Q-if/Multi-scale-Omni-Aggregation-Model.git
cd Multi-scale-Omni-Aggregation-Model

Then run inference with the downloaded Hugging Face package:

bash

PACKAGE_ROOT=/path/to/ChronusOmni-Purned-7.98b-recover

PYTHONPATH=/path/to/Chronus:$PYTHONPATH \
python scripts/inference/chronusomni_pruned_infer.py \
  --model_path "$PACKAGE_ROOT/runtime/ChronusOmni" \
  --native_pruned_ckpt "$PACKAGE_ROOT/weights/chronusomni_7p984b_teacher256_aux510_iou_grpo_2gpu_s48_ga8_lr5e7_b008.pt" \
  --json_file /path/to/VTG_dataset.json \
  --results_path ./results/submission.jsonl \
  --output_format jsonl \
  --submission_format \
  --frame_num 96 \
  --max_new_tokens 64

The output JSONL is written in compact submission format:

{"id": "...", "model_prediction": "10.03-11.80"}

Training and Recovery Code

The following code is maintained in GitHub, not as the authoritative copy in this model repository:

text

scripts/pruning/chronusomni_native_multimodal_mlp.py
scripts/training/chronusomni_native_recover_lora.py
scripts/training/chronusomni_native_grpo_recover_lora.py
scripts/inference/chronusomni_pruned_infer.py
scripts/evaluation/

Please refer to the GitHub README for pruning, SFT recovery, GRPO refinement, multi-scale inference, and evaluation commands.

Recommended File Ownership

Use this repository for:

model checkpoint files under weights/
ChronusOmni runtime files under runtime/ChronusOmni/
optional audio cache files under wavtemp/
small package-validation utilities under packaging/, scripts/packaging/, and docs/

Use GitHub for:

source code
training scripts
inference scripts
evaluation scripts
documentation of the full method

WuuuuJH/ChronusOmni-Purned-7.98b-recover

ChronusOmni 7.984B Pruned + SFT/GRPO Recovered Checkpoint

What Is Included

Expected Repository Layout

Model Checkpoint

Important Path Note

Inference

Training and Recovery Code

Recommended File Ownership

WuuuuJH/
ChronusOmni-Purned-7.98b-recover