stories15M-original
An unmodified backup of Andrej Karpathy's English stories15M.bin and its required tokenizer.
Transformer Dojo keeps this copy to preserve its original English model.
Andrej Karpathy trained the weights.
OpsBR maintains this mirror and did not train these weights.
The weights and tokenizer files are byte-identical copies of the pinned sources below.
The mirror preserves the upstream weight filename, stories15M.bin.
It contains the checkpoint that Transformer Dojo uses, without other sizes or PyTorch checkpoints from the upstream repository.
Pinned sources
The weights come from karpathy/tinyllamas.
The source revision is 0bd21da7698eaf29a0d7de3992de8a46ef624add.
UPSTREAM-README.md preserves the original model card, including its MIT metadata.
Both tokenizer files come from karpathy/llama2.c.
The source revision is 350e04fe35433e6d2941dce5a1f53308f87058eb.
They contain Meta's Llama 2 vocabulary and keep its separate license and use policy.
sources.json records the source URLs, revisions, sizes, hashes, and architecture. SHA256SUMS records the hashes of this mirror's files.
Model files
| File | Contents | Bytes | SHA-256 |
|---|---|---|---|
| stories15M.bin | Original fp32 weights in llama2.c v0 format | 60816028 | cd590644d963867a2b6e5a1107f51fad663c41d79c149fbecbbb1f95fa81f49a |
| tokenizer.bin | Original vocabulary in llama2.c format | 433869 | 50a52ef822ee9e83de5ce9d0be0a025a773d019437f58b5ff9dcafb063ece361 |
| tokenizer.model | Original vocabulary in SentencePiece format | 499723 | 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347 |
The architecture uses dim=288, hidden_dim=768, n_layers=6, n_heads=6, n_kv_heads=6, vocab_size=32000, and max_seq_len=256.
It has a tied classifier.
The weights target English story generation through llama2.c.
They are not an instruction model.
License and attribution
The model weights use the MIT license. LICENSE contains the full upstream MIT text and its copyright notice. Andrej Karpathy is the author of the weights and llama2.c.
The tokenizer files use the Llama 2 Community License Agreement and its Acceptable Use Policy. LICENSE-tokenizer.txt and USE_POLICY-tokenizer.md contain the full texts. NOTICE preserves the required Llama 2 attribution:
Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved.
The model uses the TinyStories dataset by Ronen Eldan and Yuanzhi Li. This repository contains no dataset files. The MIT license for the weights does not replace the tokenizer's license or use policy.