Downloads

Verified dataset files

Every single program in these files was compiled and executed using the real Mycel compiler (mycel v0.1.0 on Node 24). Expected outputs and compiler diagnostics are recorded from execution, not fabricated.

File	Rows	Size	Action
mycel-instruct.jsonl Instruction-tuning: beginner to app-level Mycel code	`364`	127.6 KB	Download
mycel-fix.jsonl Broken program → real compiler diagnostic → fixed program	`76`	52.7 KB	Download
mycel-translate.jsonl Python program → equivalent Mycel program (outputs match)	`40`	28.1 KB	Download
README.md (Dataset Card) Provenance, license, how it was made, format, honest limits	-	Docs	Card

Sample rows

Format preview

mycel-instruct.jsonl (sample row)

364 total rows127.6 KB

Task instructions with verified compiler output stored in meta.output:

{"messages":[{"role":"user","content":"Write a Mycel program that prints 'Hello, World!'."},{"role":"assistant","content":"print(\"Hello, World!\")"}],"meta":{"output":"Hello, World!"}}...

SHA-256: dd69584b8ded604d3660a2fbd8f2c5b9da6fc399971a9daa1446fd956245be56

mycel-fix.jsonl (sample row)

76 total rows52.7 KB

Broken code with newcomer habits, real mycel check --json diagnostic, and verified fix:

{"messages":[{"role":"user","content":"Fix this Mycel program. It fails compiler checks with:\n`def` is Python — Mycel writes `fn name(args) { ... }`\n\nBroken program:\n```mycel\ndef f(x) { }\n```"},{"role":"assistant","content":"fn f(x: Int) -> Int {\n  return x * 2\n}\nprint(f(21))"}],"meta":{"lang":"Python","habit"...

SHA-256: 23c178b1b1e138e7bbb4b85153e3467b361e6fa18a536c05bb2d7f4918482d29

mycel-translate.jsonl (sample row)

40 total rows28.1 KB

Python source to Mycel source where both programs execute and yield identical stdout:

{"messages":[{"role":"user","content":"Translate this Python program to equivalent, idiomatic Mycel:\n\n```python\nprint(\"hello, world\")\n```"},{"role":"assistant","content":"print(\"hello, world\")"}],"meta":{"python":"print(\"hello, world\")","output":"hello, world"}}...

SHA-256: 9218ce0c10cd35e36daca275a53db4609e2de846255a0baa533b9adf0a773a21

Quickstart

How to fine-tune with MLX / LoRA

These files use the standard OpenAI-compatible Chat JSONL format with messages arrays, making them compatible with MLX, HuggingFace TRL, Unsloth, Axolotl, and Ollama.

Fine-tuning on Apple Silicon (MLX LoRA)

1. Prepare training and validation splits from mycel-instruct.jsonl:

shuf mycel-instruct.jsonl > shuffled.jsonl
head -n 327 shuffled.jsonl > train.jsonl
tail -n 37 shuffled.jsonl > valid.jsonl

2. Run LoRA training with mlx_lm.lora:

python -m mlx_lm.lora \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --train \
  --data . \
  --iters 600 \
  --batch-size 4 \
  --lora-layers 16

3. Test the adapted model with a Mycel prompt:

python -m mlx_lm.generate \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --adapter-path adapters.npz \
  --prompt "Write a Mycel function that checks if an integer is prime."

Cloud Models

Teaching large cloud AI models

Looking to teach Claude, ChatGPT, or Gemini about Mycel without training weights? See /mycel/ai-learn/ for self-contained reference documents, habits tables, and system prompts you can attach or paste directly into conversation.