crypto-supreme-student-27b

Qwen3.6-27B + LoRA (r=64, all-linear), distilled from an ensemble of 6 v9 LLM teachers

• a GBM stream on enriched market features. The deployable single model of the R5 SUPREME program (June 2026).

Training

• Distill pool: 74,214 samples (phase-offset grid — provably no overlap with any teacher's train set)
• Soft labels: greedy-CV fusion (GBM out-of-fold stream selected), temperature 1.5
• Loss: 0.7·BCE(soft) + 0.3·smoothed-CE(hard, pos_weight 6), 2 epochs, bs 16, cosine LR
• This is checkpoint-7500, selected on a time-separated val window (Jan-Mar 2026) by win@10%

Results (clean v12 test, base rate 13.2%)

Model	win@top-10% conf
this model	34.3%
GBM enriched-feature baseline	34.9%
best LLM teacher (mistral_nemo)	32.1%
other v9 teachers	30.6-32.0%

Forward-looking 60-day portfolio sim (val window Jan-Mar 2026; top-5/day, +20% target, no stop, 0.2% fees): +93.2% (Sharpe 0.234) vs GBM +23.2% (Sharpe 0.090). Caveats: survivorship bias in the coin universe; midpoint-exit estimate for non-target trades; paper-trade before risking money.

Usage (scoring, NOT generation)

The training loss only constrains the Yes/No logit ratio — read token logprobs, never parse generated text:

# prob(pump) = softmax over {yes_id, no_id} of last-token logits, prompt ends with "Answer:"

Prompt format: v11 enriched (candles + "Market context:" block + "Will this coin go +15% in next 72h?").