divakantpandey commited on
Commit
9f63696
·
verified ·
1 Parent(s): 6fa122b

Add research card

Browse files
Files changed (1) hide show
  1. README.md +68 -0
README.md ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: per-source
4
+ language:
5
+ - en
6
+ - code
7
+ task_categories:
8
+ - sentence-similarity
9
+ - text-retrieval
10
+ tags:
11
+ - code-search
12
+ - code-embeddings
13
+ - hard-negatives
14
+ - negative-result
15
+ size_categories:
16
+ - 100K<n<1M
17
+ ---
18
+
19
+ # code-embed-v3-data
20
+
21
+ Replay training data for the paper *Does Execution-Verified Near-Miss Training Teach Code Embeddings to Prefer Working Code? A Negative Result* (preprint on Zenodo; DOI added on publication). Code, configs and the full research log: https://github.com/divakantpandey/code-embed-v3
22
+
23
+ These are the replay rows that both the near-miss pilot ([code-embed-v3-150m-ft-behaviour](https://huggingface.co/divakantpandey/code-embed-v3-150m-ft-behaviour)) and the replay-only control ([code-embed-v3-150m-ft-control](https://huggingface.co/divakantpandey/code-embed-v3-150m-ft-control)) trained on. The near-miss data is in [code-embed-v3-synth](https://huggingface.co/datasets/divakantpandey/code-embed-v3-synth).
24
+
25
+ ## Contents
26
+
27
+ | Folder | Columns | What it is |
28
+ |---|---|---|
29
+ | `mix_v1/` | `anchor`, `positive` | Audited pairs, up to 30,000 rows per source (`--max-rows 30000`), plus `dev/` held-out rows |
30
+ | `mix_v1_gte_mined/` | `anchor`, `positive`, `negative_1`..`negative_4` | The same pairs with four hard negatives mined by [gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base), capped at 25,000 rows per source |
31
+
32
+ Each source is a `datasets` folder saved with `save_to_disk` (Arrow). `manifest.json` lists rows, license and upstream per source; `contamination_report.json` lists every removed row.
33
+
34
+ | Source | Upstream | License | `mix_v1` rows | `mix_v1_gte_mined` rows |
35
+ |---|---|---|---|---|
36
+ | cornstack_python | nomic-ai/cornstack-python-v1 | Apache-2.0 | 29,470 | 21,468 |
37
+ | cornstack_java | nomic-ai/cornstack-java-v1 | Apache-2.0 | 29,489 | 20,246 |
38
+ | cornstack_javascript | nomic-ai/cornstack-javascript-v1 | Apache-2.0 | 29,287 | 19,643 |
39
+ | cornstack_go | nomic-ai/cornstack-go-v1 | Apache-2.0 | 29,460 | 23,966 |
40
+ | cornstack_php | nomic-ai/cornstack-php-v1 | Apache-2.0 | 29,420 | 18,831 |
41
+ | cornstack_ruby | nomic-ai/cornstack-ruby-v1 | Apache-2.0 | 28,580 | 21,315 |
42
+ | cornstack_python_c2t | nomic-ai/cornstack-python-v1 (code to text) | Apache-2.0 | 29,470 | 21,551 |
43
+ | cornstack_java_c2t | nomic-ai/cornstack-java-v1 (code to text) | Apache-2.0 | 29,489 | 18,467 |
44
+ | cornstack_javascript_c2t | nomic-ai/cornstack-javascript-v1 (code to text) | Apache-2.0 | 29,287 | 19,454 |
45
+ | code_contests_xlang | deepmind/code_contests | CC BY 4.0 | 28,966 | 649 |
46
+ | swe_smith | SWE-bench/SWE-smith | MIT | 29,500 | 23,272 |
47
+ | spider | xlangai/spider | CC BY-SA 4.0 | 6,479 | 6,463 |
48
+
49
+ The pilot and control trained on `mix_v1_gte_mined` minus rows that share an anchor with a dev row: 213,648 replay rows in total.
50
+
51
+ ## Decontamination
52
+
53
+ Every row (anchor, positive and each negative) was checked with exact and MinHash-LSH matching (5-gram, threshold 0.8) against all MTEB(Code, v1), CoIR and CoREB(v1) tasks, ExecRetrieval and the seven CoQuIR subsets. Matches were removed.
54
+
55
+ ## Loading
56
+
57
+ ```python
58
+ from huggingface_hub import snapshot_download
59
+ from datasets import load_from_disk
60
+
61
+ path = snapshot_download("divakantpandey/code-embed-v3-data", repo_type="dataset",
62
+ allow_patterns=["mix_v1_gte_mined/cornstack_python/*"])
63
+ ds = load_from_disk(f"{path}/mix_v1_gte_mined/cornstack_python")
64
+ ```
65
+
66
+ ## License
67
+
68
+ Each source keeps its upstream license (table above). The Spider rows are CC BY-SA 4.0 and must stay under that license when redistributed.