skillforge-llama-3.2-3b

SkillForge — LLaMA 3.2 3B Fine-tuned on SkillVault

A domain-specific model trained to generate professional AI agent skill files (SKILL.md) for Claude Code, Cursor, Gemini CLI, and Codex.

What It Does

Given a tool or task description, SkillForge generates a complete, structured SKILL.md file with proper YAML frontmatter, trigger conditions, prerequisites, and practical code examples.

Training

• Base model: meta-llama/Llama-3.2-3B-Instruct
• Method: LoRA fine-tuning via MLX on Apple M4
• Dataset: SkillVault — 1,113 curated SKILL.md files from trusted repos (Anthropic, LambdaTest, and community collections)
  • 10 hand-crafted synthetic examples
• Iterations: 2,000 (1,000 @ 2e-4 LR + 1,000 @ 1e-4 LR)
• Final train loss: 0.656 | Val loss: 1.312

Usage

from transformers import pipeline

pipe = pipeline("text-generation", model="ekaansh/skillforge-llama-3.2-3b")

prompt = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are SkillForge, an expert AI that generates professional agent skill files (SKILL.md).<|eot_id|>
<|start_header_id|>user<|end_header_id|>
Generate a SKILL.md for PostgreSQL database management.

Description: Best practices for queries, schemas, and debugging
Allowed tools: Bash, Read, Write<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>"""

result = pipe(prompt, max_new_tokens=600, temperature=0.7)
print(result[0]['generated_text'])

Dataset

The training dataset (SkillVault) is available at: ekaansh/skillvault