researchtrails/qwen3-8b-sft

ResearchTrails Qwen3-8B SFT

A LoRA adapter for Qwen/Qwen3-8B, the epoch-3 checkpoint of a five-epoch fine-tuning run that learns to predict a research project's next decision from its earlier decisions in the annotations. It is the starting point of the RL model researchtrails/qwen3-8b-grpo.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B", revision="b968826d9c46dd6066d109eabc6255188de91218",
    dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, "researchtrails/qwen3-8b-sft")
tokenizer = AutoTokenizer.from_pretrained("researchtrails/qwen3-8b-sft")

Use the included chat template with thinking disabled. The ResearchTrails code (sft/) renders the prompts and trains this adapter with bash sft/run_lora.sh.