babylon-mlx-engine

name	Babylon-MLX-Engine
description	C5-REAL Workspace Skill for MLX-based LLM training, evaluation, and quantization on Apple Silicon. Triggers on "pipeline", "evaluate", "loss", "MLX", "train_sft", "evalúa mis pesos", "safetensors", "Apple Silicon", "quantize", "entrenar", "pesos".

BABYLON-60 MLX ENGINE (C5-REAL SYMBIONT v2)

You are operating within the BABYLON-60 Workspace. This skill governs all interactions with the local LLM (babylon60-llm-zero).

Repository Map

babylon60-llm-zero/
├── model.py          # Transformer Decoder-Only (~120M params). RMSNorm, RoPE, SwiGLU, KV-Cache.
├── tokenizer.py      # GPT-2 BPE (tiktoken, ~50K vocab).
├── data_sft.py       # SFT Dataloader: ChatML + Asymmetric Loss Masking.
├── train_sft.py      # Legacy SFT loop (use pipeline.py instead).
├── pipeline.py       # ⭐ AUTOMATED: Train → Eval → Quantize (NaN Guard, Early Stop, Plateau Detection).
├── evaluate.py       # ⭐ 5-prompt battery test → PASS/FAIL/CATASTROPHIC_COLLAPSE verdict.
├── quantize.py       # Q4 compression (nn.QuantizedLinear, group_size=64).
├── chat.py           # Interactive REPL with streaming inference + KV-Cache.
└── *.safetensors     # Crystallized weights (BFT base / SFT fine-tuned / Q4 quantized).

Dataset: /Users/borjamoskv/.babylon60/training/datasets/train.jsonl (1,965 ChatML examples, MOSKV-1 persona).

1. Comandos Ejecutables (Scripts)

When the Operator requests training, evaluation, or pipeline execution, use these scripts directly:

Command	Script	Purpose
"Entrena el modelo" / "Lanza el pipeline"	bash .agents/skills/Babylon-MLX-Engine/scripts/run_pipeline.sh	Full Train→Eval→Quantize chain
"Evalúa mis pesos" / "Evalúa la red"	bash .agents/skills/Babylon-MLX-Engine/scripts/quick_eval.sh	5-prompt inference test + verdict

CRITICAL: Always run these via run_command with a high WaitMsBeforeAsync (send to background). Training takes minutes to hours.

2. Diagnosis Protocol

Loss Monitoring

• Use manage_task to read stdout of a running pipeline task.
• NaN detected: The pipeline auto-aborts after 3 NaN incidents. If this happens, recommend halving learning_rate in pipeline.py.
• Plateau detected: The pipeline auto-stops if loss improvement < 1% over 200 steps.
• Loss < 0.5: The pipeline auto-stops (convergence achieved).

Evaluation Verdicts

After evaluate.py runs, read eval_report.json in babylon60-llm-zero/:

• PASS: ≥3/5 coherent responses, no catastrophic repetition.
• FAIL: <3/5 coherent but no repetition collapse.
• CATASTROPHIC_COLLAPSE: Repetitive gibberish detected. Diagnose RoPE alignment or ChatML masking in data_sft.py.

Post-Mortem Actions

Verdict	Action
PASS	Proceed to quantize.py → then chat.py for interactive testing.
FAIL	Increase max_steps to 10000, or augment dataset.
CATASTROPHIC_COLLAPSE	Inspect model.py RoPE implementation. Check data_sft.py mask alignment.

3. Apple Silicon Constraints

• batch_size MUST stay ≤ 16 on M-series chips. Pipeline defaults to 8.
• Always call mx.eval() after parameter updates to flush the lazy MLX graph.
• Gradient clipping is enforced in pipeline.py (mx.clip(g, -1.0, 1.0)).

4. Zero-Fluff Rule

When reporting results, output YAML/JSON structure only. No decorative prose. Read eval_report.json or pipeline_report.json and present the data directly.