koreallmdev/deepseek70b-qlora-7816-conditional-stable-adapter

DeepSeek 70B QLoRA Adapter — 7,816 Unique Rows

Final PEFT LoRA adapter trained from deepseek-ai/DeepSeek-R1-Distill-Llama-70B.

• Unique training rows: 7,816
• Held-out Eval rows: 200
• Release label: CONDITIONAL_STABLE_WITH_QUARANTINE
• Base weights included: No
• Training dataset included: No
• Zero-defect claim: No
• Runtime repetition policy required: Yes

The repository preserves the original evaluation, known regression, quarantined cases, runtime policy, training report, and promotion record.