README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-7B-Instruct
tags:
  - video
  - forensics
  - grpo
  - temporal-grounding
pipeline_tag: video-text-to-text
---

# forensics-grpo

GRPO-trained video-forgery / temporal-forensics models, fine-tuned from
[Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct),
plus all training code, evaluation outputs and a baseline (TempSamp-R1).

Companion dataset (videos + annotations):
👉 [sdzt/forensics-grpo-data](https://huggingface.co/datasets/sdzt/forensics-grpo-data)

---

## 📂 Repository layout

```
forensics-grpo/
├── v10_r2/                     # ★ Main model
│   ├── model-0000{1..4}-of-00004.safetensors   # final weights (~16 GB)
│   ├── tokenizer / config files
│   └── checkpoint-{240,270,480,510,720,750,780,930,956}/   # 9 intermediate checkpoints
│
├── ab_noAug/                   # Ablation — no augmentation
│   ├── final weights
│   └── checkpoint-956/
│
├── ab_noHung/                  # Ablation — no Hungarian matching
│   ├── final weights
│   └── checkpoint-956/
│
├── baselines/
│   └── tempsamp_r1/            # TempSamp-R1 baseline (weights)
│       ├── TempSampR1_nocot_forensics_7B_8gpu_4ep/
│       │   ├── final weights + checkpoint-{240,480,720,956}/
│       └── TempSampR1_single_span_forensics_7B_8gpu_4ep/
│           └── final weights + checkpoint-{290,580,870,1160}/
│
├── code/                       # All source code (172 files)
│   ├── src/  scripts/          # forensics_grpo training / pipeline code
│   ├── time_r1/  tempsamp_r1/  # baseline code
│   ├── libs/  train.py         # activityforensics code
│   ├── <forensics_grpo top-level *.py>   # evaluate*.py, verifier_*.py, etc.
│   └── dl_explicit.py, dl_retry.sh       # data-download helpers
│
├── eval_results/               # Evaluation outputs (34 runs, 553 files)
│   └── eval_<run>_ckpt<N>/ ...
│
└── README.md
```

### What lives where

| Path | Contents | Size |
|------|----------|------|
| `v10_r2/` | Main model: final weights + 9 checkpoints | ~155 GB |
| `ab_noAug/` | Ablation (no aug): final + ckpt-956 | ~31 GB |
| `ab_noHung/` | Ablation (no Hungarian): final + ckpt-956 | ~31 GB |
| `baselines/tempsamp_r1/` | TempSamp-R1: 2 runs × (final + 4 ckpts) | ~155 GB |
| `code/` | Full training + eval source code | < 50 MB |
| `eval_results/` | Per-run evaluation outputs | ~19 MB |

> Each model folder holds 4-shard `model-0000x-of-00004.safetensors` weights plus
> tokenizer / processor config. `checkpoint-*` folders are weight snapshots only
> (no optimizer state).

---

## 🚀 Usage

Load the main model (final weights):

```python
from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained("sdzt/forensics-grpo", subfolder="v10_r2")
processor = AutoProcessor.from_pretrained("sdzt/forensics-grpo", subfolder="v10_r2")
```

Load a specific checkpoint or ablation:

```python
# intermediate checkpoint
model = AutoModelForImageTextToText.from_pretrained(
    "sdzt/forensics-grpo", subfolder="v10_r2/checkpoint-510")

# ablation
model = AutoModelForImageTextToText.from_pretrained(
    "sdzt/forensics-grpo", subfolder="ab_noAug")
```

Download just one folder:

```bash
hf download sdzt/forensics-grpo --include "v10_r2/*" --local-dir ./forensics-grpo
```