bonsai: depth0/experts16
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: 16 of 128 routed experts removed per layer (112 kept)
- Result: ~26.91B parameters (88.2% 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 29.
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 | 0.648 (103/159) |
90.4% of baseline accuracy at 88.2% of baseline's parameters.
Notes
Light experts-only cut: high accuracy retention but limited size savings. See depth0/experts64 for a better size/accuracy tradeoff, or depth0/experts8 for closer accuracy parity.
Project
Other checkpoints from this sweep: bonsai collection. Code and full writeup: github.com/SartajBhuvaji/bonsai (private repo).