kompress-v13 — GLM Agent Scenarios + Regex Teacher (experimental)

Training data generated by GLM-5.1-FP8 simulating realistic multi-turn coding sessions across Rust, Python, TypeScript, and Go. Labels: deterministic must-keep regex (same as production safety net). Experimental — kompress-v8 is the production recommendation.

Results

Metric	v8 (domain data)	v13 (GLM scenarios)
heretic exact (32p)	0.955	0.951
keep_rate	0.854	0.951
override_delta	0.000	0.000

Finding: GLM-generated scenarios are realistic but the regex teacher labels too many tokens as must-keep (paths, identifiers, numbers all match in code). The regex-as-teacher approach produces conservative models.

Training

127 GLM-generated tool outputs + 254 generic (33% GLM ratio). 3 epochs from v2-base.

Usage

from headroom import compress, CompressConfig
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v13"))

CONCLUSION

Regex teacher produces conservative models. 0.951 heretic but keep_rate 0.951.

USECASE

Demonstrates regex-as-teacher limitation. Educational value.

Series

Version	Teacher	Heretic	Status
v2-base	—	0.975	ceiling
v4	self-labels	0.943	—
v8	Qwen2.5-7B	0.955	production
v13	regex	0.951	too conservative

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