llm2025-basic-exp018-003

pokutuna/llm2025-basic-exp018-003

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

• Full fine-tuning (SFT) with NEFTune noise regularization
• 34,415 samples with rule-based format augmentation and camelCase key variation
• TOML/XML oversampled to address weaker performance

Training Data

All datasets are from the LLM2025 competition allowed list.

• u-10bei/StructuredOutputPractice_Train_2500
• u-10bei/StructuredOutputPractice_Train_2500_v2
• u-10bei/StructuredOutputPractice_Train_512
• u-10bei/StructuredOutputPractice_Train_512_v2
• u-10bei/StructuredOutputPractice_Train_512_v4
• u-10bei/StructuredOutputPractice_Train_512_v5

Preprocessing:

• Removed CoT (Chain-of-Thought) reasoning; train on structured output only
• Generated cross-format conversion pairs (JSON/YAML/TOML/XML/CSV)
• Added camelCase key name variants
• Fixed missing root element in XML→JSON conversions
• Filtered out rows with invalid XML tag names (numeric prefixes)
• 2x oversampling of TOML/XML to balance format distribution

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "pokutuna/llm2025-basic-exp018-003"

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)