koreallmdev/deepseek-r1-distill-llama-70b-canary1000-qlora-adapter

DeepSeek 70B Canary-1000 QLoRA Adapter

PEFT LoRA adapter only. No 70B base weights or training dataset are included. Training: 1,000 rows, one epoch, 4-bit NF4 QLoRA, seq 1024, r=8, alpha=16. Smoke20: adapter 13 wins, base 5 wins, 2 ties, 0 errors. This Canary benchmark is not a comprehensive guarantee of executable-code correctness.