code-embed-v3-data / README.md
divakantpandey's picture
Link the published preprint DOI
1d51657 verified
|
Raw History Blame Contribute Delete
3.93 kB
---
license: other
license_name: per-source
language:
- en
- code
task_categories:
- sentence-similarity
- text-retrieval
tags:
- code-search
- code-embeddings
- hard-negatives
- negative-result
size_categories:
- 100K<n<1M
---
# code-embed-v3-data
Replay training data for the paper *Does Execution-Verified Near-Miss Training Teach Code Embeddings to Prefer Working Code? A Negative Result* (preprint, [doi:10.5281/zenodo.24402953](https://doi.org/10.5281/zenodo.24402953)). Code, configs and the full research log: https://github.com/divakantpandey/code-embed-v3
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).
## Contents
| Folder | Columns | What it is |
|---|---|---|
| `mix_v1/` | `anchor`, `positive` | Audited pairs, up to 30,000 rows per source (`--max-rows 30000`), plus `dev/` held-out rows |
| `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 |
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.
| Source | Upstream | License | `mix_v1` rows | `mix_v1_gte_mined` rows |
|---|---|---|---|---|
| cornstack_python | nomic-ai/cornstack-python-v1 | Apache-2.0 | 29,470 | 21,468 |
| cornstack_java | nomic-ai/cornstack-java-v1 | Apache-2.0 | 29,489 | 20,246 |
| cornstack_javascript | nomic-ai/cornstack-javascript-v1 | Apache-2.0 | 29,287 | 19,643 |
| cornstack_go | nomic-ai/cornstack-go-v1 | Apache-2.0 | 29,460 | 23,966 |
| cornstack_php | nomic-ai/cornstack-php-v1 | Apache-2.0 | 29,420 | 18,831 |
| cornstack_ruby | nomic-ai/cornstack-ruby-v1 | Apache-2.0 | 28,580 | 21,315 |
| cornstack_python_c2t | nomic-ai/cornstack-python-v1 (code to text) | Apache-2.0 | 29,470 | 21,551 |
| cornstack_java_c2t | nomic-ai/cornstack-java-v1 (code to text) | Apache-2.0 | 29,489 | 18,467 |
| cornstack_javascript_c2t | nomic-ai/cornstack-javascript-v1 (code to text) | Apache-2.0 | 29,287 | 19,454 |
| code_contests_xlang | deepmind/code_contests | CC BY 4.0 | 28,966 | 649 |
| swe_smith | SWE-bench/SWE-smith | MIT | 29,500 | 23,272 |
| spider | xlangai/spider | CC BY-SA 4.0 | 6,479 | 6,463 |
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.
## Decontamination
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.
## Loading
```python
from huggingface_hub import snapshot_download
from datasets import load_from_disk
path = snapshot_download("divakantpandey/code-embed-v3-data", repo_type="dataset",
allow_patterns=["mix_v1_gte_mined/cornstack_python/*"])
ds = load_from_disk(f"{path}/mix_v1_gte_mined/cornstack_python")
```
## License
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.
## Citation
```bibtex
@misc{pandey2026nearmiss,
title = {Does Execution-Verified Near-Miss Training Teach Code Embeddings to Prefer Working Code? A Negative Result},
author = {Pandey, Divakant},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.24402953},
url = {https://doi.org/10.5281/zenodo.24402953}
}
```