| --- |
| license: other |
| task_categories: |
| - image-to-image |
| tags: |
| - retrieval |
| - composed-image-retrieval |
| configs: |
| - config_name: web_search |
| data_files: |
| - split: train |
| path: pairs/category=web_search/*.parquet |
| - config_name: faiss_legacy |
| data_files: |
| - split: train |
| path: pairs/category=faiss_legacy/*.parquet |
| - config_name: train-geminicot-20k |
| data_files: |
| - split: train |
| path: cot20k/pairs/category=train-geminicot-20k/*.parquet |
| - config_name: train-qwencot-20k |
| data_files: |
| - split: train |
| path: cot20k/pairs/category=train-qwencot-20k/*.parquet |
| - config_name: train-qwencot-full |
| data_files: |
| - split: train |
| path: cotfull/pairs/category=train-qwencot-full/*.parquet |
| --- |
| |
| # retdata: reasoning-based composed image retrieval pairs |
|
|
| Query image + instruction -> target image, where the target follows from the query by an |
| inference rather than visual similarity. Rows are the judge-accepted (`validity.keep`) |
| set of the general pipeline, packed in the Qwenvl2d5 v2 pair format with images inline. |
|
|
| | subset | rows | dropped (image unreadable) | |
| |---|---|---| |
| | `web_search` | 153,883 | 0 | |
| | `faiss_legacy` | 1,558,022 | 0 | |
|
|
| `pairs/category=<subset>/` holds the training rows: `query`, `cand`, `neg` structs with |
| `conversations`, `data_path` (""), `image` (`Image(decode=False)`, the original bytes) and |
| `video` (empty), as in `williamium/mmret-ume-sft-v2`. `image.path` is the source-relative |
| path (S3 `<bucket>/<key>` or `data/gsearch_targets/<file>`) for traceability; on the |
| candidate side it is the canonical path of the image's content (the smallest path among |
| targets with identical bytes), so the training collator's path-based candidate id treats a |
| target reused by several queries as one candidate. Queries use the no-think system prompt |
| and `<think></think>\n<emb>`; `neg` is null. `meta/<subset>/` has one row per training row |
| (same file name and order): the row's own source paths, blake3 content keys (`q_key`, |
| `pos_key`), instruction, caption, rationale, keywords. Some images are RGBA: decode them |
| compositing onto white, not with `.convert("RGB")` (which turns transparency black). |
|
|
| ## CoT subsets |
|
|
| `train-geminicot-20k` and `train-qwencot-20k` hold the same 20,000 rows of the two subsets above, drawn as random 100-row groups. Queries use the reasoning system prompt, with a chain of thought written by Gemini or the Qwen writer before `<emb>`. The writers saw the query image, the instruction and the target caption as a soft reference, not the target image. In `train-qwencot-20k`, 6,033 rows were written with the writer's thinking mode on (`think` in the metadata; only the final answer is kept) and 13,967 with it off. `cot20k/meta/` holds each row's source, CoT and model. Files live under `cot20k/` so that `pairs/category=*` still means the full data. |
|
|
| Rows that repeated 8 or more consecutive caption words were rewritten at a higher temperature; `temperature` in the metadata is 0 for the model default. |
|
|
| ## Full CoT subset |
|
|
| `train-qwencot-full` is the Qwen-writer CoT version of the full data, filled in as generation runs. This snapshot (2026-11-04) has 807,905 rows from 102 source shards; rows of a shard without a CoT yet are left out, and later pushes replace each file with a fuller one. Same format and prompt as `train-qwencot-20k` with the writer's thinking off; rows that repeated 8+ consecutive caption words were retried at a higher temperature (`temperature` and `caption_word_run` in `cotfull/meta/`). `model` is `writer` or `cyankiwi/Qwen3.8-27B-AWQ-INT4`, two endpoint names for the same Qwen3.8-27B-AWQ-INT4 weights. |
|
|
