kompress-v10 — Scaled C3 (experimental)

Scaled C3 self-distillation: 285 Qwen2.5-7B teacher labels + 580 generic pairs at 33% C3 ratio. Fine-tuned from chopratejas/kompress-v2-base.

More data than v8 but didn't beat it — diminishing returns on C3 labels. Base heretic 0.947, with override 0.963.

Eval

Metric	Value
heretic exact (base)	0.947
heretic exact (with override)	0.963
override_delta	+0.016
C3 pairs	285
generic pairs	580

Recommendation

Use kompress-v8 for production. v10 is here for reproducibility of the scaling experiment.

Usage

from headroom import compress, CompressConfig
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v10"))

CONCLUSION

Diminishing returns on C3 data. 285 pairs scored 0.947 vs 97 pairs at 0.955.

USECASE

Evidence that more data ≠ better. Educational value.

Full benchmark and methodology