TinyForge data

Training data for TinyForge-60M and TinyForge-60M-Instruct, a from-scratch small language model built as a personal research project. Code that produced every file here: github.com/JuaniLlaberia/tinyforge.

Contents

Path	Contents
tokenizer/vocab.json, tokenizer/merges.txt	Custom byte-level BPE tokenizer, 8,192 tokens including 4 special tokens
train/shard_0000.bin … shard_0034.bin	Pretraining tokens, uint16, 100M per shard (last shard partial): 3,470,809,194 tokens
val/shard_0000.bin	Held-out validation tokens, same format: 17,403,724 tokens
meta.json	Vocab size, special-token ids, token count per shard
stats.json	Document counts after each pipeline step
sft/train_main.jsonl	SFT set used for TinyForge-60M-Instruct: 50,031 chat conversations
sft/train_ablation.jsonl	2,031-conversation subset of the above, for the small-data ablation
sft/heldout.jsonl	1,000 held-out conversations for SFT loss
sft/stats.json	SFT filtering, truncation, token and source-mix stats

Pretraining tokens

Source: HuggingFaceFW/fineweb-edu, sample-10BT subset, ~2.9M documents. Pipeline:

Stage	Train docs	Val docs
Downloaded and split (0.5% to val by hash of doc id, before any filtering)	2,905,409	14,591
Gopher-style quality filter	2,896,126	14,543
Exact dedup	2,861,770	14,542
MinHash near-dedup	2,846,154	14,542

Dedup ran across train and val together, and whenever a duplicate spanned both splits the train copy was dropped, so no validation text leaks into training. Documents are encoded in pipeline order (not shuffled), each followed by <|endoftext|>; training draws random windows across all shards, so shard order doesn't affect it.

Shards are raw uint16 token ids: read with numpy.memmap(path, dtype=numpy.uint16, mode="r"). Special token ids: <|endoftext|> 256, <|assistant|> 257, <|user|> 258, <|endofturn|> 259 (256 raw bytes come first, so <|endoftext|> is not the last id).

SFT conversations

Source: HuggingFaceTB/smol-smoltalk (Apache-2.0), plus 31 hand-written research-style conversations (source tinyforge-research). Each line is {"messages": [{"role": "user"|"assistant", "content": ...}, ...], "source", "num_tokens", "num_trained_tokens", "num_turns"}.

• System messages are merged into the first user turn.
• Conversations longer than 1,024 tokens (in this tokenizer) are truncated at a turn boundary, keeping the longest prefix of complete user/assistant pairs that fits; dropped only if the first pair alone doesn't fit (2.7%).
• train_main is a random sample (seed 42) of smol-smoltalk's train split; train_ablation is its first 2,000 conversations plus the same 31 hand-written ones; heldout comes from smol-smoltalk's test split, with no overlap with either train set.

License

The pretraining shards and tokenizer derive from FineWeb-Edu (ODC-By 1.0); the sft/ files derive from smol-smoltalk (Apache-2.0). Use them under the terms of those sources.