tracecir / README.md
ada-flo's picture
Upload README.md with huggingface_hub
a7c3156 verified
|
Raw
History Blame Contribute Delete
2.68 kB
# TraceCIR / RevCIR
A **multi-turn revision-layer** Composed Image Retrieval (CIR) benchmark over fashion.
Each chain is a shopper iteratively editing an image search (POS / NEG / MIXED /
COMMIT→RETRACT / cross-category turns), with per-turn query forms (rule tag-diff,
FashionIQ-style terse, and a persona-styled rich utterance).
## Download (latest version)
Data is hosted on HuggingFace (private): **[`fashiontrace/tracecir`](https://huggingface.co/datasets/fashiontrace/tracecir)**
(request access; then `huggingface-cli login`).
Current renderer version = **v4** (multi-turn rebuild, all chains ≥3 turns). The
`releases/latest/` path always points at the version the viewer currently renders:
```bash
# full set / splits (jsonl, one chain per line)
huggingface-cli download fashiontrace/tracecir releases/latest/tracecir.jsonl --repo-type dataset --local-dir .
huggingface-cli download fashiontrace/tracecir releases/latest/tracecir.val.jsonl --repo-type dataset --local-dir .
huggingface-cli download fashiontrace/tracecir releases/latest/tracecir.test.jsonl --repo-type dataset --local-dir .
```
Direct URL (needs auth for the private repo):
`https://huggingface.co/datasets/fashiontrace/tracecir/resolve/main/releases/latest/tracecir.jsonl`
Images are a single shared pool, archived as **`images_pool.tar`** in the repo root
(~3.59 GB, **142,704 images**; covers every release version v1–v4). This full pool is the
retrieval gallery for eval; only 9,768 distinct images appear in the 2,988 chains, so the
gallery is ~15× the referenced set (what makes Recall@k meaningful). Download + extract once:
```bash
huggingface-cli download fashiontrace/tracecir images_pool.tar --repo-type dataset --local-dir .
tar xf images_pool.tar # -> images/<category>/<image_id>.jpg
```
Image id in the jsonl = the filename stem (e.g. `fashion200k_dress_00123` or a FashionIQ id).
## Versions
- **v4** (`releases/v4/`, current): multi-turn rebuild — 2988 chains, category-balanced
(498/cat×6), **all ≥3 turns (~82% ≥4 turns)**, persona-styled rich queries per turn.
Judged consistently by **Qwen3.6-35B-A3B (thinking)**; failed turns recovered by
re-judge + salvage (image-grounded edit re-derivation). val 1470 / test 1518.
- v3 = `data/revcir_balanced.json` (2982 chains, 75% 2-turn). v1/v2 text = `versions/`.
## Code / render
- Chain generation, judging, salvage, rebuild: `src/` (see `docs/handoff/HANDOFF.md`).
- HTML viewers: `src/render_diverse.py`, `scripts/build_compare_viewers.py`
→ `sample/revcir_v4.html` (and v1–v3).
- Data version map: `data/DATA_VERSIONS.md`. Env recreation: `docs/handoff/requirements_venv*.txt` (uv).