kompress-v14 — GLM Council Training Loop (experimental)

First kompress model trained with a GLM-5.1-FP8 "council" controlling the training loop. After each training round, GLM reviews metrics (loss, heretic, keep_rate) and decides: SHIP or RETRAIN with more epochs. Experimental proof-of-concept — kompress-v8 is the production recommendation.

Council Decisions

Round	Epochs	Council	Final Heretic
1	3	RETRAIN	—
2	5	RETRAIN	—
3	7	RETRAIN	0.882

Finding: The council correctly identified underperformance (always RETRAIN) but more epochs caused overfitting. The council concept works but 3 epochs remains the empirical sweet spot regardless of what an LLM says.

Usage

from headroom import compress, CompressConfig
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v14"))

CONCLUSION

Council training concept proven. GLM-5.1 correctly identified underperformance but more epochs caused overfitting.

USECASE

Proof of concept for LLM-controlled training loops. Not for production.

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