AXE-cue — the 0.6B cue chain of Peabirux

Two LoRA adapters over Qwen/Qwen3-0.6B (revision c1899de289a04d12100db370d81485cdf75e47ca), trained with mlx_lm.lora on Apple silicon. Together they form the distilled cue chain reported in §6.7 of Verified XPath synthesis with open-weight proposers (WebConf 2027 submission): a DOM pruner followed by a cue extractor that reads the signalling label and the value of each requested field from a pruned page. They are a cost study, not a column of the proposer matrix.

folder	role	adapters.safetensors sha256
teacher_pruner_v4_qwen3_0_6b/	DOM pruner (keeps the subtrees around every field occurrence)	64489d8f2d2d71922ffa3d8ddb9154a545fe5b2a69ff2b8a0014ee2cceebfd10
cue_extractor_v1/	cue extractor (label + value per field on the pruned DOM)	d4a2350ddbdcc69adfb64d390a8b27fbab8319220ccc80af6c8020dac6fce590

Each folder carries the mlx_lm adapter_config.json, the generated LoRA config, the training log and a checkpoint_manifest.json (base revision, prompt-template hash, data manifest). Absolute paths of the training workstation are replaced by <repo> / <home>.

Training recipe (both): LoRA rank 64, alpha 64, dropout 0.1, AdamW lr 1.7e-5 with cosine decay and 10% warmup, weight decay 0.01, batch 1, gradient checkpointing, max sequence 4096 tokens, 2 epochs, seed 20260825, prompt masked.

See docs/REPRODUCING.md (run type axe-retrain) in the repository for the registered numbers and the exact protocol.