This repository mirrors the contents of the project's local data/ folder — the dataset, not the codebase. To use it inside the project, copy this repo's files into:

data/

Snapshot was cleaned & restructured on 2026-07. The catalog/graph audit numbers below reflect the 2026-06 tagging review (the most recent full re-tag).

Scope — what belongs vs. out of scope

In this dataset repo	NOT here (kept local only — gitignored)
`custom/` catalog, graph KB, legacy outfits, body images	Stylist model checkpoints / LoRA adapters
`cache/store_registry.db` + `cache/raw/<store>/` (scrape provenance)	`cache/hf/`, `cache/turboquant/` downloaded models
`reports/` quality audits (CSV/JSON/MD)	Model-eval & inference scratch (`cache/eval`, `cache/probe_.jsonl`, `cache/tagging_.log`)
`stylist/` conversation training corpus (selected JSONL/CSV)	Kaggle run infra, wheelhouses, raw training logs

Removed on 2026-07 cleanup: stale empty skeletons raw/ (except raw/occasion_cache/*.db, kept on-disk as a local tagger cache, no longer synced), processed/, the unused DVC pointer raw.dvc, and the throwaway tmp_kaggle_output_probe/.

Current validated snapshot (2026-06 re-tag)

Catalog

5618 catalog items
5618 item-store links
0 invalid rows in the latest item-tag audit
0 high-severity semantic flags in the latest item-tag audit
556 items have empty desc_vi (accepted warning)
Categorization rule: store-native product_type first, title-token fallback
Semantic tagging coverage: 5618 / 5618 non-empty rows
Color tagging coverage: 5618 / 5618 non-empty rows
Main-garment semantic completeness (top | bottom | dress | outerwear): 5004 / 5004

Store distribution:

Store	Items
`yody_vn`	2301
`aristino_vn`	1606
`canifa_vn`	684
`rubies`	468
`huelleyrose`	328
`dirtycoins`	231

Category distribution:

Category	Items
`top`	2800
`bottom`	1457
`outerwear`	383
`dress`	364
`accessory`	348
`bag`	136
`shoes`	130

Gender distribution:

Gender	Items
`men`	2549
`women`	1873
`kid`	696
`unisex`	500

Graph KB (primary retrieval artifact)

4694 adult item nodes (men | women | unisex)
316559 canonical compatibility edges in custom/graph/item_edges.parquet
Latest audit/eval status:
- coherence_violations = 0
- fitb_recall@5(mask shoes) = 1.0000
- Shoeless outfits are treated as valid clothing cores; missing shoes are tracked as a completion gap rather than an invalid core combo.

Legacy materialized outfits

custom/outfits/generated_outfits.parquet is retained only for legacy comparison/debug.
The project's primary KB path is graph retrieval, not materialized outfits.
Latest outfit-tag audit over assembled graph outfits:
- total_outfits = 824
- valid_core_outfit_count = 824
- complete_outfit_count = 816
- shoeless_valid_core_count = 8
- high_severity_flag_count = 0

Layout

cache/
  store_registry.db          # scrape registry (store id -> adapter) — enables rescrape
  raw/<store_id>/             # raw HTTP/HTML scrape cache per store (reproducibility)
custom/
  catalog/
    catalog_metadata.parquet
    item_store_links.parquet
    scrape_manifest.json
    images/                   # 5618 item thumbnails
  graph/
    item_edges.parquet        # primary compatibility graph KB
  outfits/
    generated_outfits.parquet   # legacy/comparison only
  body/
    images/                   # (placeholder) body-condition images
reports/                     # quality audit reports (full set incl. CSV)
  tagging_quality/            # item-tag audit (CSV/MD/JSON)
  outfit_tagging_quality/     # outfit-tag audit (CSV/JSON)
stylist/                     # stylist conversation training corpus (selected JSONL/CSV)
  fine_tune/
    runs/...
    stylist_knowledge/

Minimum files for graph retrieval / eval

custom/catalog/catalog_metadata.parquet
custom/catalog/item_store_links.parquet
custom/catalog/images/
custom/graph/item_edges.parquet

cache/store_registry.db + cache/raw/<store>/ are optional but recommended for reproducible rescrape. reports/ and stylist/ are documentation/training artifacts and are not required for retrieval.

Stylist conversation corpus (`stylist/`)

Conversation data used to LoRA-fine-tune the Qwen3-VL stylist. The heavy Kaggle run infra, checkpoints, and wheelhouses are local-only and intentionally not synced here; only the actual SFT corpus is published.

Distilled fine-tune dataset (canonical, retrieval-grounded)

stylist/fine_tune/runs/stylist_grounded_v2/
  final_bundle_gptoss_2800_merged_core8800/
    train.jsonl     # 11256 examples (~14.5 MB)
    eval.jsonl      #   344 examples
    manifest.json   # task/source counts
  merged_source_gptoss_2800_core8800/
    manifest.json   # records the 2 source files merged

This is the final SFT bundle. It merges a retrieval-grounded GPT-OSS-teacher distillation (gptoss_2800_all_accepted.jsonl, 2800 grounded dialogues) with the knowledge core (stylist_distilled_qwen35_under10k/train.jsonl, 8800 rows) into 11600 total examples stratified to 11256 train / 344 eval. Task breakdown (from manifest.json): 7200 stylist_knowledge + 2800 grounded_generated (900 tool_calling_grounded, 700 recommend_explain_grounded, 350 ask_missing_info_grounded, 300 body_fit_grounded, 250 no_result_or_relax, 150 polite_decline_anti_hallucination, 150 multi_turn_grounded) + 1600 behavioral_synthetic (tool_calling, ask_missing_info, recommend_explain, multi_turn, polite_decline, body_analysis, edge_case). Every tool-call row uses the canonical <tool_call>...<tool_call> / ... wire format locked in src/outfitmatch/stylist/tools.py (enum args from vocab.py).

Earlier distilled corpora (kept for traceability)

stylist_distilled_qwen35_under10k/ — pre-ground distilled corpus (~10k knowledge rows) reused as the knowledge core above.
stylist_distilled_behavioral/ — behavioral / tool-call synthetic seeds.
kaggle_qlora_token3_t4_qwen/kaggle_dataset/ — packing set used for one Kaggle QLoRA run.
stylist_knowledge/finetuning_data_fashion_knowledge.csv — raw knowledge source table.

Reports / quality audits (`reports/`)

Two audit categories, both generated from the catalog/graph snapshot:

reports/tagging_quality/ — item-level tag distribution, invalid rows, outliers, semantic flags, plus model-comparison review markdown files for the tagging tagger selection (Gemma 4 12B TurboQUANT vs Qwen3-VL 8B vs Qwen3.5 9B).
reports/outfit_tagging_quality/ — outfit-level tag distributions by core shape / gender / gen method / price tier, invalid rows, semantic flags, and the summary counts quoted in the snapshot above.

These are present here so the dataset card is self-documenting; they are also the artifacts referenced by docs/EXPERIMENT_GUIDE.md for the 2026-06 tagging review.

Schema highlights

`custom/catalog/catalog_metadata.parquet`

Column	Meaning
`item_id`	Stable item id: `item_custom_NNNNN`
`category`	`top`, `bottom`, `dress`, `outerwear`, `shoes`, `bag`, `accessory`
`source_product_type`	Original store taxonomy value before mapping
`gender`	`men`, `women`, `unisex`, `kid`
`formality`	`athletic`, `casual`, `smart_casual`, `formal`
`image_path`	Relative image path
`title_vi`	Product title
`desc_vi`	Product description
`colors`	JSON string list
`collected_date`	Collection date
`collector`	Collector/job name

`custom/catalog/item_store_links.parquet`

Column	Meaning
`item_id`	FK to catalog
`store_id`	Store identifier
`source_product_id`	Raw source product id
`product_url`	Purchase URL
`price_vnd`	Normalized VND price
`sale_price_vnd`	Optional sale price
`sku`	Optional SKU
`in_stock`	Stock flag
`available_sizes`	JSON string list
`sizes_in_stock`	JSON string list

`custom/graph/item_edges.parquet`

Column	Meaning
`src_id`	Source item id
`dst_id`	Destination item id
`src_category`	Source item category
`dst_category`	Destination item category
`weight`	Compatibility weight

Notes

Out-of-scope SKUs have been removed: underwear, swimwear, phone cases, perfume, gift vouchers, keychains, 2-piece sets.
source_product_type is persisted for traceability.
The project code that consumes this dataset lives in a separate repository: fashion_match_project.

Downloads last month: 13

Total file size:

4.24 GB

You need to agree to share your contact information to access this dataset

VN Fashion Data