---
tags:
- code
- pretraining
- with-in-us-ai
---

# AGA-Code Pretrain Corpus v1.0.0

Clean, deduplicated pretraining corpus for AGA-Code (Autonomous God Agents),
built from six WithinUsAI coder datasets.

## Sources

| Source | Raw | Unique | Notes |
|---|---|---|---|
| WithinUsAI/GOD_Coder_100k | 100,000 | 100,000 | file was gzip mislabeled as .jsonl; decompressed |
| WithinUsAI/Elite_GOD_Coder_100k | 100,000 | 100,000 | same gzip repair |
| WithinUsAI/Omega_Genesis_Coder_100k | 100,000 | 100,000 | same gzip repair |
| WithinUsAI/python_GOD_coder_100k | 100,000 | 99,498 | same gzip repair |
| WithinUsAI/GOD_Coder_Complete_DataSet | 175,000 | 173,860 | `messages` chat column rendered as `role: content` blocks |
| WithinUsAI/Royal_Ghost_Coder_501k | 490,000 | 422,124 | instruction+response; ~14% internal duplicates removed |

## Build

- Tolerant JSONL parsing (bad lines counted, none found after gzip repair).
- Records under 64 chars dropped.
- **SHA-256 dedup** on normalized text: 1,065,000 raw → **995,482 unique** (69,518 removed).
- **Seed-1147 shuffle**, then split: **993,482 train** / **2,000 validation** (held-out).
- Schema: `{"text": ..., "source": ...}` — one JSON per line.

## Files

- `train.jsonl` (993,482 rows)
- `validation.jsonl` (2,000 rows, held-out)
- `manifest.json` (per-source counts, SHA-256 of each file, seed)

(c) 2026 WithIn Us AI — Guy DuGan II