Qwen3-1.7B — AIME SC DPO
Standalone model produced by merging the aime_dpo_qwen3_1_7b_sc_lora_fullcoverage LoRA adapter into Qwen/Qwen3-1.7B.
The merge was performed in float32 and exported as bfloat16 safetensors. The tokenizer and chat template are included. Loading does not require a separate adapter or PEFT.
Use AutoModelForCausalLM.from_pretrained and AutoTokenizer.from_pretrained with this repository ID. Apply the included Qwen chat template when formatting conversations.
Merge provenance and numerical validation are recorded in merge_info.json. This upload does not establish downstream AIME accuracy.