Data selection is usually studied for efficiency. DRIFT asks whether it can also raise the ceiling and produce a better model than training on all of the data.
DRIFT starts from a model that has finished SFT on its full corpus. It samples the model's own responses to validation queries, signs them by correctness, and ranks the original corpus by its influence on those responses. Training then continues on the top 10%. Influence functions are a local approximation, and the model's own responses make much closer targets than external reference answers.
Results
Accuracy (%) after continued SFT on 10% of each model's original SFT corpus. Small numbers give the change from the full-data checkpoint.
| Method | Olmo3-7B-Instruct-SFT | OpenR1-Distill-7B | ||
|---|---|---|---|---|
| All | Add. | All | Add. | |
| Full data | 39.17 | 25.49 | 57.23 | 46.24 |
| Random | 39.04−0.13 | 25.21−0.28 | 57.40+0.17 | 46.65+0.41 |
| DSIR | 34.37−4.80 | 22.76−2.73 | 56.24−0.99 | 45.43−0.81 |
| QuRating | 39.26+0.09 | 25.71+0.22 | 57.78+0.55 | 47.05+0.81 |
| BM25 | 39.54+0.37 | 25.92+0.43 | 57.32+0.09 | 46.54+0.30 |
| LESS | 37.63−1.54 | 23.96−1.53 | 57.28+0.05 | 46.44+0.20 |
| RDS | 39.08−0.09 | 25.50+0.01 | 57.63+0.40 | 46.91+0.67 |
| IF (GraSS) | 38.66−0.51 | 24.80−0.69 | 57.85+0.62 | 47.69+1.45 |
| DRIFT | 40.14+0.97 | 26.51+1.02 | 58.66+1.43 | 48.73+2.49 |
Full data is the model after its original SFT on the whole corpus. Every other row continues from it. All averages eight benchmarks, and Add. averages the four that are never used for attribution.
Release
- CodeSelection pipeline, training recipe and evaluation harness
- Selected subsetThe 192,453 conversations DRIFT selected for Olmo3
- Validation rollouts8,220 queries and 263,040 sampled responses with correctness labels
- Per-run results57 continued-SFT runs behind the main table and the ablations