# Time-R1 baseline on ActivityForensics

This directory only adds data adapters; **all upstream Time-R1 code is unchanged**
(`finetune.py`, `src/time_r1/*`, etc.). Reward (single-span IoU + optional format)
is taken as-is from the original repo.

## What gets adapted

ActivityForensics annotations are multi-segment (a video has K ≥ 1 forged
intervals). Time-R1 only models single-span temporal grounding, so each
`(video, segment)` pair is emitted as one training example with a fixed query:

> "an AI-manipulated segment"

Multi-segment GT is split per segment; all splits of the same video share the
same Qwen2.5-VL preprocessed cache via symlinks.

## Setup

```bash
cd /mnt/local-fast/zhangt/baselines/time_r1

# 1) Build Time-R1-format annotation JSON.
python data_forensics/build_forensics_json.py \
    --annot_dir /ces/zt/activityforensics/annot \
    --video_root /ces/zt \
    --output_dir ./dataset/forensics/annotation
# -> writes train.json / val.json

# 2) Symlink the pre-encoded video tensors that forensics_grpo already produced.
python data_forensics/link_cache.py \
    --forensics_cache /mnt/local-fast/zhangt/forensics_grpo/<your_cache_dir> \
    --annotation_json_dir ./dataset/forensics/annotation \
    --output_dir ./dataset/forensics/preprocessed
# -> creates ./dataset/forensics/preprocessed/{train,eval}/<vid>_<sid>/...

# 3) Launch training (edit MODEL_PATH at the top of the script first).
bash scripts/finetune/run_forensics.sh
```

`build_forensics_json.py` uses forensics_grpo's `data_loader` parser, so any
change to its annotation schema flows through automatically. Override
`FORENSICS_GRPO_ROOT=<path>` if forensics_grpo lives elsewhere.

## Files added

- `data_forensics/build_forensics_json.py`  - .txt → Time-R1 .json
- `data_forensics/link_cache.py`            - cache layout symlinker
- `scripts/finetune/run_forensics.sh`       - launch script