kompress-v6

Token compression classifier fine-tuned from PeetPedro/kompress-v4 (ModernBERT-base, 149M params). Trained as part of the ultrawhale fine-tuning loop.

Kompress classifies each token in a message as keep (1) or drop (0). Used by the headroom proxy to compress LLM context before it reaches the model.

Eval results (heretic adversarial benchmark)

Heretic-style prompts generate responses maximally dense with must-keep tokens (chemical formulas, CVE identifiers, memory addresses, line numbers). The benchmark measures what fraction of those tokens survive compression.

Metric	Value
heretic exact_pct	0.962
keep_rate	0.854
override_delta	0.000
base model	kompress-v4

Full progression across all versions

Training

3,000 synthetic Claude Code agent-pattern pairs (bash_output, file_read, error_trace, search_result, json_tool_result) merged with 2,003 generic pairs. Fine-tuned from v4. Self-labeling skipped — v4 subword tokenizer degrades agent references (mk_in_ref collapsed to 0.652). Generator references (mk_in_ref=1.0) used directly. Model became more conservative (keep_rate 0.854 vs v4's 0.823).

Usage

# Via headroom proxy (recommended)
# ANTHROPIC_BASE_URL=http://localhost:8787 claude

# Direct library use
from headroom import compress, CompressConfig
result = compress(messages, config=CompressConfig(kompress_model="PeetPedro/kompress-v6"))

Series

Version	heretic	keep_rate	Notes
v3	0.942	0.728	first self-label
v3.1	0.925	—	domain data
v3.2	0.929	—	domain refined
v3.3	0.942	—	domain-only, overfit
v4	0.967	0.823	override internalized
v5	0.961	—	loop converged
v6	0.962	0.854	agent-distribution

Training code: ultrawhale