pokutuna/llm2025-main

Fine-tuned model based on Qwen/Qwen3-4B-Instruct-2507 for structured output generation (JSON / YAML / XML / TOML / CSV).

This repository contains model weights and tokenizer.

Training Objective

Improve structured output accuracy for the StructEval-T benchmark.

• Loss is applied only to the assistant output (assistant_only_loss); intermediate Chain-of-Thought reasoning is masked
• NEFTune noise regularization for generalization

Training Data

All datasets are from the LLM2025 competition allowed list. Preprocessed with rule-based structuring and filtering; a subset is used for training.

Dataset	License
u-10bei/structured_data_with_cot_dataset_512_v2	MIT
u-10bei/structured_data_with_cot_dataset_512_v4	MIT
u-10bei/structured_data_with_cot_dataset_512_v5	MIT
u-10bei/structured_data_with_cot_dataset_512	not specified
u-10bei/structured_data_with_cot_dataset_v2	not specified
u-10bei/structured_data_with_cot_dataset	not specified
daichira/structured-3k-mix-sft	CC-BY-4.0
daichira/structured-5k-mix-sft	CC-BY-4.0
daichira/structured-hard-sft-4k	CC-BY-4.0
u-10bei/dpo-dataset-qwen-cot	not specified

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "pokutuna/llm2025-main"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Convert the following text to JSON: ..."}]
text = tokenizer.apply_chat_template(
  messages,
  tokenize=False,
  add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=1024, do_sample=False)

generated = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))

License

• Model: Apache-2.0 (inherited from base model)
• Training Data: MIT / CC-BY-4.0 (see Training Data section)