Hardtests Sft Qwen3.5 27b R16 Epoch1

--- base_model: Qwen/Qwen3.5-27B library_name: peft tags: - lora - sft - qwen3.5 - hardtests - competitive-programming --- # HardTests SFT — Qwen3.5-27B LoRA (epoch 1) LoRA adapter from supervised fine-tuning of `Qwen/Qwen3.5-27B` on the [ZhuofengLi/hardtests-sft-qwen3.5-27b](https://huggingface.co/datasets/ZhuofengLi/hardtests-sft-qwen3.5-27b) dataset (reward=1 competitive-programming trajectories with reasoning). - **Checkpoint**: step 100 (recorded epoch = 1.1249112845990064) - **LoRA**: r=8, alpha=32 (see `adapter_config.json`) - **Contents**: adapter weights + configs only (no optimizer state) ## Merge into the base model ```python from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-27B", torch_dtype="bfloat16", device_map="auto") model = PeftModel.from_pretrained(base, "ZhuofengLi/hardtests-sft-qwen3.5-27b-r16-epoch1") model = model.merge_and_unload() model.save_pretrained("./qwen3.5-27b-hardtests-sft-epoch1-merged") AutoTokenizer.from_pretrained("Qwen/Qwen3.5-27B").save_pretrained("./qwen3.5-27b-hardtests-sft-epoch1-merged") ```

Source context: 8 downloads · 0 likes · Library peft · Repo ZhuofengLi/hardtests-sft-qwen3.5-27b-r16-epoch1