merlin-corpus-v0 — v0.0.1

Tokenized pretraining corpus for Merlin, a 3B LLM for agentic coding on Apple Silicon.

Note: v0 does not yet include agentic traces. Tokenizer and corpus will be retrained after trace generation.

Contents

• corpus_train.bin + corpus_val.bin, uint16, shape [N, 6144]
• 90/10 train/val split at document boundaries (seed=42 shuffle)
• ~1.19B tokens total, 100% packing efficiency
• Fits entirely in H100 HBM (~2.4GB)
• Tokenizer: tsuberim/merlin-tokenizer-v0

Sources

Source	Records	Notes
The Stack — Python	~730K	AST-parsed, has def/class, no generated files
The Stack — Bash	~214K	≥2 shell commands
The Stack — Markdown	~63K	English, has backtick, no HTML blobs
Stack Exchange Q&A	~21K	Python/bash/linux/git/docker topic filter
GitHub commits	~84K	Non-trivial commit messages
GitHub issues	~71K	Topic-filtered, no bot posts
tldr-pages	~7K	Clean command references

Format

import numpy as np
# Load into memory (~2.4GB total, fits in H100 HBM)
train = np.fromfile("corpus_train.bin", dtype=np.uint16).reshape(-1, 6144)  # (N, 6144)
val   = np.fromfile("corpus_val.bin",   dtype=np.uint16).reshape(-1, 6144)