v8_stage1_json_csv-merged

Model Description

This model is Stage 1 of the Sequential Format Learning (v8 strategy) for structured data output.

Training Strategy

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

• Train one format at a time
• Merge LoRA to base model after each stage
• Use merged model as the base for the next stage

Stage 1 Focus: JSON & CSV

• Formats: JSON (400 samples) + CSV (400 samples) = 800 total
• Goal: 100% parse success rate for JSON and CSV
• Base Model: Qwen/Qwen3-4B-Instruct-2507

Training Parameters

• MAX_SEQ_LEN: 1024
• EPOCHS: 2
• Learning Rate: 5e-05
• LoRA R: 64, Alpha: 128

Sequential Format Learning Pipeline

Stage 1: JSON/CSV (800) ← This model
    ↓
Stage 2: YAML (500)
    ↓
Stage 3: XML (500)
    ↓
Stage 4: Mixed/TOML (1000)
    ↓
Final Model → LB 0.8+

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("kmd2525/v8_stage1_json_csv-merged")
tokenizer = AutoTokenizer.from_pretrained("kmd2525/v8_stage1_json_csv-merged")

Next Stage

Use this model as the base for Stage 2 (YAML training):

os.environ["SFT_BASE_MODEL"] = "kmd2525/v8_stage1_json_csv-merged"