MoECat-LLM

---

Overview

MoECat is an end-to-end reasoning distillation pipeline for Scratch 3.0 algorithm education. The project converts 3K+ raw .sb3 algorithm works into standard scratchblocks pseudocode, then distills them via a top-tier reasoning LLM into high-quality JSONL corpora with <think> tags and deep algorithmic breakdowns, and finally fine-tunes lightweight MoE/LLM models under Kaggle's free dual-T4 GPU constraints.

Pipeline

.sb3 Raw Files
    |
    v
[1] Node.js AST Converter       (.sb3 --> scratchblocks pseudocode)
    |
    v
[2] LLM Distillation Skill      (single-item / read-and-destroy --> JSONL with <think>)
    |
    v
[3] Kaggle Fine-tuning          (Unsloth 4-bit QLoRA, dual T4)
    |
    v
MoECat Series Models

Stage	Tool	Output
Scratch Parsing	Node.js + sb3 parser	scratchblocks text
CoT Distillation	DeepSeek-R1-Distill-Qwen-8B	JSONL with <think> reasoning
Fine-tuning	Unsloth + QLoRA 4-bit	LoRA adapter weights

Target Model

DeepSeek-R1-Distill-Qwen-8B

Project Structure

MoECat-LLM/
├── dataset_pipeline/
│   ├── scratch_scraper.py           # Scratch project batch scraper (auto-converts to scratchblocks)
│   ├── parse-sb3-blocks/            # Node.js sb3-to-scratchblocks converter
│   │   ├── cli.js                   # CLI entry point
│   │   ├── src/                     # Parser, block mapping, sanitizer source
│   │   ├── rollup.config.js         # Build config
│   │   └── package.json
│   └── Compact-Reasoning-Distiller/ # Agent skill for CoT distillation
│       ├── SKILL.md                 # Skill definition and workflow
│       ├── get_unprocessed_id.py    # Tool: get next unprocessed project ID
│       └── merge_dataset.py         # Tool: merge all entries into final JSONL
├── fine_tune/                       # (planned) Unsloth QLoRA training scripts
├── examples/                        # (planned) Advanced Scratch algorithm examples
├── LICENSE
└── README.md

License

Apache-2.0