kompress-v12 — Qwen3-Coder-Next Teacher (experimental)

C3 self-distillation using Qwen3-Coder-Next as the labeling teacher. Experimental — kompress-v8 is the production recommendation.

Results

Metric	v8 (Qwen2.5-7B)	v12 (Qwen3-Coder)
heretic exact (32p)	0.955	0.949
keep_rate	0.854	0.949
override_delta	0.000	+0.001

Finding: Qwen3-Coder-Next is biased toward preserving ALL code tokens — great for coding, counterproductive for compression. The weaker Qwen2.5-7B teacher found a better keep/drop balance.

Training

141 Qwen3-Coder-Next labeled pairs + 282 generic (33% C3 ratio). 3 epochs from v2-base, batch 16, lr 2e-5.

Usage

from headroom import compress, CompressConfig
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v12"))

CONCLUSION

Qwen3-Coder teacher too conservative. 0.949 heretic. Stronger teacher ≠ better compression teacher.

USECASE

Shows teacher bias problem. Use Qwen2.5-7B instead.

Series

Version	Teacher	Heretic	Status
v2-base	—	0.975	precision ceiling
v4	self-labels	0.943	override internalized
v8	Qwen2.5-7B	0.955	production
v12	Qwen3-Coder	0.949	too conservative

Full benchmark → | Training repo → | Headroom → | vaked.dev →