Simplify dataset card

README.md

@@ -3,176 +3,30 @@ language:
 - en
 task_categories:
 - text-retrieval
-- question-answering
 tags:
 - rag
 - retrieval
-- query-encoder
 - reler
 - flashrag
 - qwen3-embedding
-- nq
-- hotpotqa
-pretty_name: RELER RAG (G3) Data
-size_categories:
-- 10M<n<100M
 license: other
 ---
 
-# RELER RAG (G3) Data
+# RELER RAG Data
 
-This repository contains the **data used to train and evaluate the RAG query-encoder
-experiments (G3)** in the RELER project. It is a self-contained snapshot: frozen
-Wikipedia corpus, the frozen E0 FAISS index, FlashRAG training/evaluation queries,
-mined candidate lists, passage-level qrels, and evaluation outputs.
+Data for the RELER RAG query-encoder experiments (G3).
 
-> **No model weights or training checkpoints are included.** Only data and
-> derived data artifacts.
+| Path | Contents |
+|---|---|
+| `rag/corpus/` | `wiki18_100w`: 21,015,324 Wikipedia passages |
+| `rag/qwen3_e0_index/` | Frozen Qwen3-Embedding-0.6B FAISS index (1024-d, fp16, 64 shards) |
+| `rag/flashrag/` | NQ/HotpotQA train queries + 7-dataset eval queries |
+| `rag/raw/` | DPR `biencoder-nq-train.json.gz` (NQ labels) |
+| `rag/candidates/` | Top-1000 candidates for 108,730 training queries |
+| `rag/qrels*/` | Binary passage qrels (NQ: DPR score==1000; HotpotQA: supporting facts) |
+| `rag/eval/` | Evaluation outputs |
 
-## 仓库内容 / Contents
+Upstream sources: `RUC-NLPIR/FlashRAG_datasets` @ `bcafb8dd`, DPR `biencoder-nq-train.json.gz`,
+`Qwen/Qwen3-Embedding-0.6B` @ `97b0c61`. Upstream licenses apply.
 
-```
-rag/
-├── corpus/
-│   ├── wiki18_100w.jsonl              # 21,015,324 Wikipedia passages (13.4 GB)
-│   └── wiki18_100w.offsets.npy        # byte offsets for O(1) passage lookup
-├── qwen3_e0_index/                    # frozen E0 FAISS index (41 GB, 64 shards)
-│   ├── index_manifest.json            # hashes, model revision, pooling protocol
-│   ├── vectors-00000-of-00064.npy     # float16, 1024-d, L2-normalized
-│   └── ...
-├── flashrag/                          # FlashRAG_datasets @ bcafb8dd (701 MB)
-│   ├── nq/{train,test}.jsonl
-│   ├── hotpotqa/{train,dev}.jsonl
-│   ├── triviaqa/test.jsonl
-│   ├── popqa/test.jsonl
-│   ├── 2wikimultihopqa/dev.jsonl
-│   ├── musique/dev.jsonl
-│   └── bamboogle/test.jsonl
-├── raw/
-│   └── biencoder-nq-train.json.gz     # DPR NQ train positives (2.2 GB)
-├── candidates/
-│   ├── nq_hotpotqa_train.jsonl        # top-1000 candidates, 108,730 queries (2.8 GB)
-│   ├── nq_hotpotqa_train.jsonl.manifest.json
-│   └── anchors/                       # fp16 E0 query embeddings (anchor regularization)
-├── qrels/
-│   ├── nq_hotpotqa_train.jsonl        # official binary passage qrels
-│   ├── nq_hotpotqa_train.jsonl.manifest.json
-│   └── nq_hotpotqa_train.unresolved.jsonl
-├── qrels_f1_0.5/                      # relaxed HotpotQA mapping (ablation)
-├── qrels_f1_0.8/                      # identical to qrels/ (kept as named ablation)
-├── qrels_alignment_cache/             # intermediate mapping cache
-├── eval/                              # evaluation outputs (7 QA datasets)
-│   ├── g3_e0/                         # untrained E0 baseline
-│   ├── g3_cl_s42/, g3_retrl_mrr_s42/, g3_retrl_ndcg_s42/
-├── generator_cache.sqlite3            # frozen-generator replay cache
-├── flashrag_manifest.json             # upstream provenance + per-file SHA-256
-└── question_leakage.json              # 60 normalized train/eval question overlaps
-```
-
-## 数据来源与版本 / Provenance
-
-| Artifact | Upstream | Version / hash |
-|---|---|---|
-| Retrieval corpus | [FlashRAG `wiki18_100w`](https://huggingface.co/datasets/RUC-NLPIR/FlashRAG_datasets) | revision `bcafb8dd07d453be3cbeeeb3f78be1841bddf92c`, archive SHA-256 `e5b987a0…b956db` |
-| QA train/eval queries | same FlashRAG snapshot | see `rag/flashrag_manifest.json` |
-| NQ passage labels | [DPR `biencoder-nq-train.json.gz`](https://dl.fbaipublicfiles.com/dpr/data/retriever/biencoder-nq-train.json.gz) | SHA-256 `32492315…fe40c73` |
-| HotpotQA passage labels | FlashRAG `hotpotqa/train.jsonl` supporting facts | SHA-256 `a81274ab…64d1a` |
-| E0 document encoder | [Qwen/Qwen3-Embedding-0.6B](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B) | revision `97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3` |
-| Frozen generator (answer rewards) | [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) | not included; run separately |
-
-### Training queries
-
-- Sources: NQ `train` (79,168 rows) + HotpotQA `train` (90,447 rows) = 169,615.
-- After requiring complete passage-label mapping: **108,730 queries**
-  (NQ 58,879; HotpotQA 49,851). HotpotQA queries whose supporting facts could
-  not be mapped are dropped (40,547); NQ drops 1.
-- Default config holds out a deterministic 5% tuning split
-  (`sha256("20260909:{source}:{query_id}")` bucket) → 103,293 train / 5,437 tuning.
-  `configs/experiments/iclr2027/g3.yaml` uses `rag_split: full` (all 108,730).
-
-### Labels
-
-- **NQ**: DPR `positive_ctxs` with `score == 1000` only.
-- **HotpotQA**: all annotated supporting facts mapped by title + sentence.
-- `qrels_f1_0.5/` relaxes the mapping with a title + window token-F1 ≥ 0.5 rule
-  (66,223 HotpotQA queries kept); `qrels_f1_0.8/` is byte-identical to `qrels/`
-  (SHA-256 `7ad52b9a…05144d`) and is kept only as the named ablation artifact.
-- Candidates carry three judgments per passage: `training_positive_mask` (qrels),
-  `answer_positive_mask` (golden answer string occurs in passage), and
-  `evidence_group_ids` (multi-hop evidence groups, used by graded relevance).
-- All known positives are forced into each depth-1000 list; unjudged documents
-  are treated as zero gain.
-
-### Corpus and index
-
-- 21,015,324 English Wikipedia passages (100-word passages, 2018 dump).
-- Encoded with Qwen3-Embedding-0.6B: last-token pooling, left padding, `pad`
-  append token, max 512 tokens, float16, L2-normalized, 1024 dimensions,
-  64 shards. Built once and frozen; RAG training only updates the query encoder.
-
-### Evaluation suite (not used for training)
-
-51,713 queries over 7 datasets; in-domain = NQ/HotpotQA, held-out = the other 5:
-
-| Dataset | Split | Queries |
-|---|---|---|
-| NQ | test | 3,610 |
-| TriviaQA | test | 11,313 |
-| PopQA | test | 14,267 |
-| HotpotQA | dev | 7,405 |
-| 2WikiMultiHopQA | dev | 12,576 |
-| MuSiQue | dev | 2,417 |
-| Bamboogle | test | 125 |
-
-## 使用方式 / Usage
-
-Download into a checkout of RELER so that paths resolve to `data/rag/...`:
-
-```bash
-hf download liuqi6777/RELER-RAG-data --repo-type dataset --local-dir data
-# => data/rag/corpus, data/rag/qwen3_e0_index, ...
-```
-
-Then, from the RELER repo root:
-
-```bash
-# Train (4 GPUs); see configs/rag/*.yaml
-scripts/rag_pipeline.sh train configs/rag/rl_mrr.yaml
-
-# Evaluate a checkpoint on the fixed 7-dataset suite
-scripts/rag_pipeline.sh eval --model <checkpoint> --output-dir data/rag/eval/rerun
-```
-
-To rebuild artifacts from the upstream sources instead of downloading them:
-
-```bash
-scripts/rag_pipeline.sh prepare      # download + verify FlashRAG / wiki18_100w
-scripts/rag_pipeline.sh encode       # 4-GPU encode -> qwen3_e0_index
-scripts/rag_pipeline.sh qrels        # build passage qrels
-scripts/rag_pipeline.sh candidates   # mine top-1000 candidates
-```
-
-All artifacts are hash-pinned in their `*.manifest.json`. Verification is on by
-default in the pipeline; pass `--skip-hash-verification` only if you intentionally
-modified an input.
-
-## 已知问题 / Known issue
-
-`rag/question_leakage.json` records **60 normalized-question overlaps** between
-the NQ training queries and NQ test queries (upstream duplication). They are
-documented rather than removed; filter by `query_id` if you need a strict split.
-
-## 许可 / License
-
-The derived artifacts (qrels, candidate lists, index, evaluation outputs) are
-released for research use. **Upstream data retains its original license and
-terms**; please follow them when redistributing:
-
-- FlashRAG datasets: https://huggingface.co/datasets/RUC-NLPIR/FlashRAG_datasets
-- Natural Questions: https://ai.google.com/research/NaturalQuestions
-- HotpotQA: https://hotpotqa.github.io/
-- Wikipedia text: CC BY-SA
-- Qwen3-Embedding-0.6B / Qwen2.5-7B-Instruct: Apache-2.0
-
-If you are a rights holder and believe an artifact should not be redistributed,
-please open a discussion on this repository.
+File sizes are listed in `MANIFEST.tsv`.