File size: 2,133 Bytes
c716592
 
 
 
 
 
 
 
 
 
 
 
 
 
74ff54c
c716592
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
74ff54c
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
---
license: other
license_name: per-source
language:
- en
- code
tags:
- code-search
- code-embeddings
- negative-result
---

# ce3-ablate-data

Small-scale training mixes 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

| Archive | Scale | Used by |
|---|---|---|
| `dryrun_ablate_x10.tar` | Up to 3,000 replay rows per source, 400 CodeContests problems, 2,756 SWE-smith twins; 29,939 train rows, 11% near-miss | Three-seed comparison (paper Table 1, log R9) and batch-size bridge (Table 2, log R15) via `experiments/ablate6.py` and `experiments/bridge.py` |
| `dryrun_ablate_r3.tar` | 1x scale (about 300 rows per source) | Earlier ablation rounds (`experiments/ablate3.py`, `ablate4.py`; log R5, R7). R5 ran before the float16 master-weight fix (log R6) and is not valid evidence |

Each archive extracts to `dryrun_ablate/` with `data/` (pairs), `behaviour/` (near-miss build), `combined/` (mined replay plus near-miss rows, as trained) and `configs/`. Folders are `datasets` saved with `save_to_disk`.

```bash
tar -xf dryrun_ablate_x10.tar
```

## Decontamination

These mixes were audited against two MTEB tasks before training. A full audit afterwards (log R13) found no overlap with ExecRetrieval or CoQuIR Defects4J; 15 rows (0.05%) overlap CoQuIR CodeNet, which no small-scale result uses.

## License

Rows keep their upstream licenses: CoRNStack Apache-2.0, CodeContests CC BY 4.0, SWE-smith MIT, Spider CC BY-SA 4.0. Generated rows (mutants, rewrites, `synth_*`) come from Qwen2.5-Coder-7B-Instruct (Apache-2.0).

## 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}
}
```