RinnRinnmini/lora_structeval_t_qwen3_4b_sft_v2

This repository provides a LoRA adapter (PEFT) fine-tuned from Qwen/Qwen3-4B-Instruct-2507 for structured output tasks (StructEval-T style).

Important: This repo contains LoRA adapter weights only. Load the base model separately, then apply this adapter.

Training Objective

Improve structured output accuracy and stability for formats such as JSON / YAML / XML / TOML / CSV.

Training Data

• (not specified in metadata) Set SFT_DATASET_ID env var to record the dataset id(s).

Usage (example)

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

base_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "RinnRinnmini/lora_structeval_t_qwen3_4b_sft_v2"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

prompt = "Convert this JSON to YAML. Return ONLY YAML.\n\nJSON:\n{\"a\": 1, \"b\": [2,3]}"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

License / Terms

• Base model: Qwen/Qwen3-4B-Instruct-2507 (see base model license/terms)
• This repo: LoRA adapter weights + tokenizer-related files.