qwen3-4b-structeval-T-20251230-lora

This repository provides a LoRA adapter fine-tuned on Qwen3-4B-Instruct-2507 for structured output generation and format conversion, optimized for the StructEval-T benchmark.

---

Model Overview

• Base Model: unsloth/Qwen3-4B-Instruct-2507
• Fine-tuning Method: QLoRA (via Unsloth)
• Adapter Type: LoRA (PEFT)
• Target Benchmark: StructEval-T
• Language: Japanese / English mixed (instruction-level)

This adapter is designed to improve the model’s ability to:

• Generate pure structured outputs (JSON / YAML / TOML / XML / CSV)
• Perform format-to-format conversions (e.g., JSON → YAML, XML → JSON)
• Strictly follow output-only constraints (no explanations, no markdown fences)
• Produce parser-valid and structurally consistent outputs

---

Training Data

Dataset Characteristics

The model was fine-tuned on a custom SFT dataset specifically constructed for StructEval-T, with the following properties:

• Task Families

  • generation: generate structured output from instructions
  • conversion: convert between structured formats

• Supported Formats

  • JSON
  • YAML
  • TOML
  • XML
  • CSV

• Design Principles

  • Only samples where the user explicitly requests a structured format are included
  • Assistant outputs are strictly output-only
  • All samples are parser-validated and structure-checked
  • Bucket-aware deduplication applied to maximize learning efficiency
  • Minimum coverage ensured for each important (task × format) bucket

Dataset Scale (after deduplication)

• Total samples: ~14k
• Overall duplication rate: ~0.13
• All major StructEval-T buckets: ≥ 500 samples
• JSON / YAML / TOML buckets: near-zero duplication
• XML / CSV buckets: limited diversity but sufficient for syntax learning

---

Training Configuration

• Fine-tuning Framework: Unsloth
• Quantization: QLoRA (4-bit base weights)
• Optimizer: Paged AdamW (32-bit)
• Learning Rate: 2e-4
• Batch Size: 16 (global)
• Epochs: 1

The training setup prioritizes instruction-following accuracy and format correctness over general language modeling.

---

Intended Use

This LoRA adapter is intended for:

• Evaluating and improving StructEval-T performance
• Structured output generation pipelines
• Format conversion tasks in downstream agents or tools
• Research and educational experiments on instruction-following for structured data

Not Recommended For

• Open-ended conversational agents
• Creative text generation
• Long-form reasoning or chain-of-thought tasks

---

Usage Notes

When using this adapter, prompts should:

• Explicitly specify the target output format
• Request output-only structured data
• Avoid ambiguous or conversational instructions

This matches the conditions under which the adapter was trained.

---

License

This adapter follows the license of the base model:

• Apache-2.0 (same as Qwen3-4B-Instruct-2507)

---

Acknowledgements

• Qwen / Alibaba Cloud for the base model
• Unsloth for the efficient QLoRA training framework
• StructEval benchmark authors for the evaluation methodology