README.md

metadata

license: apache-2.0
base_model: kmd2525/v8_stage3_xml-merged
tags:
  - structured-output
  - json
  - csv
  - yaml
  - xml
  - toml
  - sft
  - sequential-format-learning
language:
  - en
  - ja

v8 Final Model - Sequential Format Learning

Model Description

This is the final model of the Sequential Format Learning (v8 strategy) for structured data output. Trained through 4 stages to achieve optimal performance across all structured formats.

Training Strategy

Based on Person U's approach that achieved 0.84 on the leaderboard:

• Train one format at a time, building format-specific expertise progressively
• Merge LoRA to base model after each stage
• TOML learning comes from other formats (Person T discovery)

Completed Training Stages

Stage	Formats	Samples	Target
Stage 1	JSON, CSV	800	100% parse success
Stage 2	YAML	500	100% parse success
Stage 3	XML	500	95%+ parse success
Stage 4	Mixed (all formats)	1000	TOML 90%+, overall optimization

Stage 4 (Final) Details

• Base Model: kmd2525/v8_stage3_xml-merged (Stage 3 merged model)
• Dataset: 1000 samples, balanced across all 5 formats (200 each)
• Goal: Final tuning for TOML and overall balance

Training Parameters

• MAX_SEQ_LEN: 1024
• EPOCHS: 2
• Learning Rate: 2e-05 (lower for final stage)
• LoRA R: 64, Alpha: 128

Expected Performance

Format	Target
JSON	100%
CSV	100%
YAML	100%
XML	95%+
TOML	90%+

Target LB Score: 0.8+

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("kmd2525/v8_final_model")
tokenizer = AutoTokenizer.from_pretrained("kmd2525/v8_final_model")

# Example inference
messages = [
    {"role": "user", "content": "Convert the following to JSON: name=Alice, age=30"}
]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Acknowledgments

This model was developed based on insights from the StructEval competition community, particularly Person U's sequential format learning approach and Person T's discovery about TOML learning.