Hyukkyu commited on
Commit
298799e
·
verified ·
1 Parent(s): 24bb44e

eli5: hard negatives and teacher scores (jina35)

Browse files
README.md CHANGED
@@ -52,9 +52,8 @@ in this collection.
52
  | Score values | 1 ×325,473 |
53
  | Layout | `queries` · `corpus` · `qrels` · `hard-negatives` · `teacher-scores`, split `train` |
54
  | Splits | `corpus`: train · `hard-negatives`: train · `qrels`: train · `queries`: train · `teacher-scores`: train |
55
- | Hard negatives | **none yet** — config present with 0 rows |
56
- | Teacher scores | **none yet** — config present with 0 rows |
57
- | Reading empty configs | `datasets` cannot return a 0-example split (`load_dataset` raises "corresponds to no data"); until rows exist, read the Parquet directly with `pyarrow`/`polars`/`pandas`. The schema is declared in the file and in `configs:` above |
58
  | Ids | `sha1(text)[:20]`; identical texts collapse to one document (182 collapsed) |
59
  | Pair recovery | 325,473 of 325,473 source pairs reconstructed from `queries` × `qrels` × `corpus` with byte-equal text |
60
  | License | `unspecified` |
@@ -76,6 +75,29 @@ before publishing; `provenance.json` records the source file hashes, what change
76
 
77
  - **byte-preserved** all text — no whitespace, newline, or control-character normalisation
78
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
79
  ## Load it
80
 
81
  ```python
 
52
  | Score values | 1 ×325,473 |
53
  | Layout | `queries` · `corpus` · `qrels` · `hard-negatives` · `teacher-scores`, split `train` |
54
  | Splits | `corpus`: train · `hard-negatives`: train · `qrels`: train · `queries`: train · `teacher-scores`: train |
55
+ | Hard negatives | sources: `dense` · 9,928,863 rows |
56
+ | Teacher scores | `jinaai/jina-reranker-v3.5` · 10,028,863 rows (positives included) |
 
57
  | Ids | `sha1(text)[:20]`; identical texts collapse to one document (182 collapsed) |
58
  | Pair recovery | 325,473 of 325,473 source pairs reconstructed from `queries` × `qrels` × `corpus` with byte-equal text |
59
  | License | `unspecified` |
 
75
 
76
  - **byte-preserved** all text — no whitespace, newline, or control-character normalisation
77
 
78
+ ## Hard negatives and teacher scores
79
+
80
+ Filled by the owner's annotation pipeline (`annotation=jina35`) for the `train` split of the
81
+ query set(s) below; queries without a labelled positive are left out.
82
+
83
+ - **Candidates**: dense retrieval with `jinaai/jina-embeddings-v5-text-small` over the full corpus to depth 1,000;
84
+ 100 candidates per query drawn from the rank windows 1–30 (30), 31–100 (30), 101–300 (20), 301–1000 (20), the query's labelled positives excluded.
85
+ `rank` is the dense rank; `source` is `dense` for a mined row and `dataset` for a negative the source labels itself
86
+ (those are kept for every query of the split, sampled or not).
87
+ - **Teacher**: `jinaai/jina-reranker-v3.5`, listwise: a query's positive and all of its candidates are scored together in one
88
+ context of up to 32,768 tokens. `score` is the raw cosine score, one row per (query, positive) and per
89
+ (query, candidate); a labelled negative that was also mined is scored once. No filtering is applied to the tables.
90
+
91
+ | configs | queries | hard negatives | teacher scores |
92
+ |---|---:|---:|---:|
93
+ | `hard-negatives` · `teacher-scores` | 100,000 (seeded sample, seed 1) | 9,928,863 (9,928,863 dense) | 10,028,863 |
94
+
95
+ ```python
96
+ from datasets import load_dataset
97
+ negatives = load_dataset("Hyukkyu/train-eli5", "hard-negatives", split="train")
98
+ scores = load_dataset("Hyukkyu/train-eli5", "teacher-scores", split="train")
99
+ ```
100
+
101
  ## Load it
102
 
103
  ```python
hard-negatives/{train-00000-of-00001.parquet → train-00000-of-00002.parquet} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f62b1b56789db979f3b0d83e5e329c7d6533a72d3da081396053e4bb2a4543b0
3
- size 823
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:596554fc1b072e62a4f6912810de3e1656cab9bbb487adf696e5ac46f3d94169
3
+ size 60336626
teacher-scores/train-00000-of-00001.parquet → hard-negatives/train-00001-of-00002.parquet RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8316266e7cbe6867227eefc104eb93ee2252b885a6127a24b8e5c1982716406a
3
- size 815
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a94fb9b0207f676abfd7c25588886d36ca37ceca767ac8a5a4eeb09127f99cc8
3
+ size 60312689
teacher-scores/train-00000-of-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:adbf4f1f6a829f8607f77e9e6e25e0a65819e9c2b3cc3b4069cc260ac6edc28e
3
+ size 76020740
teacher-scores/train-00001-of-00002.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9a57fccd682f6d080629456c1a01efa5b26bf704540de90eb2dff394c6caa85e
3
+ size 76020815