🧠 Cognitive Twin — Training Runbook

# 1. Merge all SFT data
cd Desktop/Comp-Core/packages/cognitive-twin
python3 scripts/local_finetune.py

# Output: local_finetune/data/train.jsonl, valid.jsonl, test.jsonl
# Stats: ~16K deduplicated records from 41K raw

{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

# 1. Ensure MLX is installed on Mac4
ssh mac4 'pip3 install mlx mlx-lm'

# 2. Rsync data
rsync -avz Desktop/Comp-Core/packages/cognitive-twin/local_finetune/ mac4:[home-path]

# 3. Train
ssh mac4 'bash -lc "
export PATH=\$HOME/Library/Python/3.9/bin:\$PATH
mlx_lm.lora \
  --model mlx-community/gemma-3-1b-it-4bit \
  --data [home-path] \
  --adapter-path [home-path] \
  --train \
  --batch-size 1 \
  --num-layers 4 \
  --iters 500 \
  --learning-rate 5e-5 \
  --max-seq-length 256 \
  --steps-per-report 10 \
  --grad-checkpoint
"'

# 4. Test inference
ssh mac4 'bash -lc "
export PATH=\$HOME/Library/Python/3.9/bin:\$PATH
mlx_lm.generate \
  --model mlx-community/gemma-3-1b-it-4bit \
  --adapter-path [home-path] \
  --max-tokens 200 \
  --prompt \"What projects are you working on?\"
"'

# 5. Backup adapters
rsync -avz mac4:[home-path] Desktop/Comp-Core/packages/cognitive-twin/local_finetune/adapters/

1. Go to https://colab.research.google.com
2. Sign in as [email]
3. File → Upload notebook → select notebooks/twin_finetune_colab.ipynb
   OR create new notebook and upload train_twin.py
4. Runtime → Change runtime type → A100 GPU
5. Upload files:
   - train.jsonl (192MB)
   - valid.jsonl (10MB)
   - train_twin.py
6. Run cells in order, or just:
   !pip install -q unsloth trl peft accelerate bitsandbytes
   !python train_twin.py
7. Download twin-lora-adapter.zip when done

export TOGETHER_API_KEY=$(grep TOGETHER [home-path] | cut -d= -f2)

# 1. Upload data (if not already uploaded)
python3 -c "
from together import Together
import httpx
client = Together([sensitive field redacted], timeout=httpx.Timeout(600.0))
train = client.files.upload(file='$HOME/Desktop/Comp-Core/packages/cognitive-twin/local_finetune/data/train.jsonl', purpose='fine-tune')
val = client.files.upload(file='$HOME/Desktop/Comp-Core/packages/cognitive-twin/local_finetune/data/valid.jsonl', purpose='fine-tune')
print(f'Train: {train.id}')
print(f'Val: {val.id}')
"

# 2. Launch fine-tuning job
curl -X POST https://api.together.xyz/v1/fine-tunes \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-Next-80B-A3B-Instruct",
    "training_file": "<TRAIN_FILE_ID>",
    "validation_file": "<VAL_FILE_ID>",
    "suffix": "twin-alpha-v2",
    "n_epochs": 2,
    "n_evals": 10,
    "n_checkpoints": 1,
    "batch_size": 16,
    "learning_rate": 2e-5,
    "training_type": {
      "type": "Lora",
      "lora_r": 16,
      "lora_alpha": 32
    }
  }'

# 3. Monitor
curl https://api.together.xyz/v1/fine-tunes/<JOB_ID> \
  -H "Authorization: Bearer $TOGETHER_API_KEY" | python3 -m json.tool

# 4. After completion — deploy as serverless LoRA
# The model_output_name from the job response is your endpoint

test_prompts = [
    "What projects are you currently working on?",
    "How would you approach building a new iOS app?",
    "Explain the N'Ko keyboard architecture",
    "What's the status of the BWB POS app?",
    "How does the Dream Garden work?",
    # ... add domain-specific prompts
]

cognitive-twin/
├── scripts/
│   ├── local_finetune.py          # Data merge + MLX prep
│   ├── density_v4.py              # Density scoring
│   └── gen_v9_*.py                # SFT data generators
├── notebooks/
│   ├── twin_finetune_colab.ipynb  # Interactive Colab notebook
│   └── train_twin.py              # Standalone Colab script
├── colab-upload/                   # Pre-packaged for Colab
│   ├── train.jsonl
│   ├── valid.jsonl
│   └── train_twin.py
├── local_finetune/
│   ├── data/                      # Merged, deduplicated data
│   ├── adapters/                  # Mac4 local LoRA adapters
│   └── manifest.json
├── data/
│   ├── ctv3_export_v3/            # Base corpus (41K records)
│   └── expansion_v9/             # V9 generators output
├── MAC4_BENCHMARK.md              # Local model benchmark results
└── TRAINING_RUNBOOK.md            # This file