v8_curated_1k-merged

Model Description

This model is trained using the Person W Style approach - high-quality data under 1000 samples.

Training Strategy (Person W Style)

Based on Person W's insight that achieved 0.8+ with:

• High-quality data under 1000 samples
• epoch=2
• Fine-tuning hyperparameters rather than large changes

Dataset: v8_curated_1k

Quality Criteria:

• 100% parse success rate
• No code fence contamination
• No natural language prefix/suffix
• Appropriate complexity

Format Distribution:

• Balanced across all 5 formats (JSON/CSV/YAML/XML/TOML)

Training Parameters

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

Comparison with Sequential Format Learning

Approach	Data	Training	Complexity
Sequential SFT (Stage 1-4)	2800 samples	4 stages	High
Person W Style (This model)	< 1000 samples	1 stage	Simple

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("kmd2525/v8_curated_1k-merged")
tokenizer = AutoTokenizer.from_pretrained("kmd2525/v8_curated_1k-merged")