CardCauldronTCG/basil-glimmer-qlora

Basil (Glimmer base) — Card Cauldron TCG label-skeptic adapter

A QLoRA adapter for meta-models/Muse-Glimmer-30B (pinned revision a4e59da52a7bc87ae7251dd5545c0dd437c44b68), fine-tuned by Card Cauldron TCG for the Meta / Cerebral Valley hackathon (submission deadline Nov 4, 2026). Basil reads a PROPOSED trading-card label plus the crew's standing rules and answers with one JSON object: AGREE, or QUESTION with the rule and the missing piece. He never gives a real/fake verdict, a grade, or a price.

Training

• Data: the frozen "true-v2 feast" — 5,893 rows, SHA-256 7a01050d9093ac119bb44aef40b1921dd7e1159ada0e324aced2fa30618f9fd2 (identity-disjoint from the held-out exam split by construction).
• Method: QLoRA (4-bit NF4 base, bf16 compute) per Meta's official cookbook recipe; adapters on language projections only, vision tower frozen. LoRA r16 / alpha 32 / dropout 0.05, lr 1e-4 cosine, 3 epochs, effective batch 32, max seq 1024, seed 29, completion-only loss. Trained on Modal (1x H100).
• This upload is checkpoint checkpoint-185.

Held-out exam (§5, 305 rows never trained on)

Bars: decision >= 0.90 · citation >= 0.85 · forbidden outputs = 0 · UNKNOWN-preservation violations = 0.

• decision accuracy: 1.0
• rule citation accuracy: 1.0
• forbidden outputs: 0
• UNKNOWN-preservation violations: 0
• passed all four bars: True

Where it runs

Live endpoint (Modal): https://caudillsabastian6--glimmer-basil-serve-basil-read.modal.run The submitted runtime serves this adapter on the pinned base above; nothing else is in the runtime's model path.

Licence

Apache-2.0, inheriting the base model's licence.