Jyler/KnightHacks-Laya-Model

KnightHacks Laya Model

Laya SPOT triage models from the run4 LoRA sweep.

Laya (421M, ModernBERT-large) fine-tuned on the SPOT triage data with amd/system1/scripts/sweep_lora.sh (12 epochs max, patience 4, checkpoint selected on calibration cross-entropy). Trained on an RTX 3080 Ti.

Each run folder has best/ (lowest calibration CE) and final/ (last epoch), both loadable with laya.load(path) (LoRA merged into model.safetensors; LoRA runs also include lora_adapter/), plus training_report.md with every epoch.

Results at the saved best/ checkpoint (mean macro-F1; untuned Laya: real 0.426)

Run	Strategy	Best epoch / run	Calib CE	Real	Synthetic	Counterfactual	Overall
run4_A_lora	LoRA r16, lora lr 2e-4	4 / 8	0.314	0.788	0.782	0.962	0.784
run4_B_lora_lr4	LoRA r16, lora lr 4e-4	6 / 10	0.258	0.802	0.786	0.979	0.797
run4_C_full	full fine-tune, enc lr 2e-5, llrd 0.95	4 / 8	0.313	0.816	0.796	0.964	0.803
run4_D_lora_r64	LoRA r64/alpha128, lora lr 2e-4	4 / 8	0.265	0.802	0.826	0.971	0.811
run4_E_loraplus	LoRA+ x16, lora lr 1e-4	5 / 9	0.279	0.803	0.781	0.969	0.799

Recommended: run4_B_lora_lr4/best (lowest calibration CE). Real-segment differences between runs are within the noise of 9 held-out missions.

Download

hf download Jyler/KnightHacks-Laya-Model --local-dir KnightHacks-Laya-Model                      # everything (~8.5 GB)
hf download Jyler/KnightHacks-Laya-Model --include "run4_B_lora_lr4/best/*" --local-dir models     # one checkpoint