qwen3-4b-structeval-T-20251231_v3-full_lora

This repository provides a LoRA adapter fine-tuned on Qwen3-4B-Instruct-2507 for structured output generation and format conversion, specifically optimized for the StructEval-T benchmark (v3 Strategy).

---

Model Overview

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

This v3 model is designed to overcome the limitations observed in previous versions, particularly:

• XML / YAML / TOML support: Drastic improvement in non-JSON/CSV structural outputs.
• Overfitting Mitigation: Balanced CSV-to-JSON training to handle both hierarchical and flat structures.
• Instruction Consistency: Normalized prompt format aligned with the StructEval-T benchmark.

---

Training Data (v3 Hybrid Mix)

The model was fine-tuned on the structeval_t_sft_final_v3_normalized.jsonl dataset, a curated mix of the original v1 and the expanded v2 datasets.

• Total samples: ~2,000
• Sampling Strategy:
  • v1 Selection (1,000 samples): Smart undersampling of the 1,800-sample v1 dataset to cap over-represented hierarchical CSV tasks.
  • v2 Expansion (1,000 samples): Targeted generation of low-scoring task formats (XML, TOML, YAML) and robust tabular data.

Task Categories & Composition

Category	Task Classes	Description	Samples (est)
Advanced Formats	C4, C5, C6, G4-8	XML (with attributes), YAML, TOML generation.	~700
Tabular Robustness	C1 (Flat/Hier)	Balanced CSV to JSON conversion (fixing hierarchical overfitting).	~550
Core Conversion	C2, C3, G1	Standard JSON/CSV/XML conversions and entity extraction.	~750

Key Improvements in v3

• Format Expansion: Includes training on XML attributes (<item id="123">) and complex nested YAML/TOML structures.
• Hierarchical Neutralization: Significant reduction in "hierarchical-only" CSV bias from v1, improving performance on standard/flat CSV files.
• Seed Diversification: Integrated seeds from libpostal, wikitext, openfoodfacts, and foursquare via streaming sampling.

---

Training Configuration

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

---

Intended Use

• StructEval-T Evaluation: Specifically tuned for high performance on structural consistency and format conversion tasks.
• Data Engineering Agents: Reliable extraction and conversion of various structured data formats.

---

License

This adapter follows the license of the base model:

• Apache-2.0

Acknowledgements

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