bonsai: depth0/experts60
Pruned and distilled from Qwen3-30B-A3B-Base for TypeScript code completion. Part of the bonsai project, which applies the Minitron structured-pruning and knowledge-distillation recipe (Muralidharan et al. 2024, Sreenivas et al. 2024) to a Mixture-of-Experts model, a pruning axis neither paper covers.
What was pruned
- Depth: 0 of 48 layers removed (48 kept)
- Experts: 60 of 128 routed experts removed per layer (68 kept)
- Result: ~16.94B parameters (55.5% of baseline's ~30.5B)
Training
500-step fixed-loss short schedule (75-step warmup, peak LR 1e-4), distilled against the uncorrected teacher. Best-validation checkpoint, step 49 (final-step checkpoint overfits: val loss 0.204, not used here).
Results
HumanEval-TS (TypeScript MultiPL-E), full 159-problem suite, temperature 0.2, top-p 0.95:
| pass@1 | |
|---|---|
| Qwen3-30B-A3B-Base (unpruned) | 0.717 (114/159) |
| This model | 0.572 (91/159) |
79.8% of baseline accuracy at 55.5% of baseline's parameters.
Notes
Nearly identical size and accuracy to depth0/experts64 (from an earlier batch in the same sweep). Use depth0/experts64 unless there's a specific reason to prefer this one.
Project
Other checkpoints from this sweep: bonsai collection. Code and full writeup: github.com/SartajBhuvaji/bonsai (private repo).