Back to corpus
technical noteexperiment writeup candidatescore 24

🧠 Cognitive Twin — Training Runbook

| Platform | Model | Cost | Time | Quality | |----------|-------|------|------|---------| | **Mac4 Local** | gemma-3-1b-it (4-bit) | $0 | 5 min | ⭐⭐ (proof of concept) | | **Google Colab Pro** | gemma-3-12b-it (4-bit) | $0 (subscription) | 1-2h | ⭐⭐⭐⭐ | | **Together AI** | Qwen3-Next-80B-A3B | ~$16-20 | 2-4h | ⭐⭐⭐⭐⭐ | | **Together AI** | Qwen3-235B-A22B | ~$100-200 | 8-12h | ⭐⭐⭐⭐⭐⭐ (future) |

Open full HTMLDownload HTMLCorpus JSONPaper schema

Full HTML reader

Read the full artifact

Open in new tab

Extracted abstract or opening context

> Repeatable pipeline for fine-tuning the Cognitive Twin on any platform. > Last updated: 2026-02-18 | Platform | Model | Cost | Time | Quality | |----------|-------|------|------|---------| | **Mac4 Local** | gemma-3-1b-it (4-bit) | $0 | 5 min | ⭐⭐ (proof of concept) | | **Google Colab Pro** | gemma-3-12b-it (4-bit) | $0 (subscription) | 1-2h | ⭐⭐⭐⭐ | | **Together AI** | Qwen3-Next-80B-A3B | ~$16-20 | 2-4h | ⭐⭐⭐⭐⭐ | | **Together AI** | Qwen3-235B-A22B | ~$100-200 | 8-12h | ⭐⭐⭐⭐⭐⭐ (future) | ### Data Location - **Train:** `Desktop/Comp-Core/packages/cognitive-twin/local_finetune/data/train.jsonl` (192MB, 16,360 records) - **Val:** `Desktop/Comp-Core/packages/cognitive-twin/local_finetune/data/valid.jsonl` (10MB, 909 records) - **Test:** `Desktop/Comp-Core/packages/cognitive-twin/local_finetune/data/test.jsonl` (10MB, 909 records) **⚠️ Gemma models require strict alternation:** `system? (user assistant)+` The merge script handles this automatically. ### When to use - Quick iteration, testing data quality - Zero cost, ~5 minutes - Limited to 1B model (16GB RAM constraint)

Promotion decision

What has to happen next

Attach run IDs, datasets, metrics, and reproduction commands.

Why this is not always a full paper yet

Corpus pages are public-safe readers for discovered workspace artifacts. They are not automatically final papers. A corpus item becomes a polished paper only after the editable source, evidence checkpoints, references, figures, render path, and release status are attached through the paper schema.