Modelpublic

mila-llm/gpt2-small

sourceHugging Facemitupdated 7d agoView on Hugging Face

0likes

Model Card

GPT-2 Small — FP32 for Mila

openai-community/gpt2, converted for Mila and published in safetensors format, at the precision the weights were released in.

622 MiB. It is the reference model in Mila's catalogue rather than a deployment target: the smallest thing that exercises the whole path end to end, and the only published model that is not quantized.

What this is for

GPT-2 is a base model — it continues a prompt. It was not instruction-tuned, so it reads no chat template, holds no turns, and calls no tools. Three consequences worth stating up front, because each of them is a design decision rather than a defect:

—The Mila chat harness refuses it, keyed on the manifest's instruct: false. Chat renders turns and applies a template; a base model uses none of that.
—The Mila Inference Server does not serve it. MIS speaks the OpenAI and Anthropic chat protocols, which are instruct-shaped.
—Completion is the mode it belongs to, along with the training path — Mila's Bard sample trains a GPT-2 stack from scratch.

It is also the one to reach for when what you are testing is Mila rather than the model: 622 MiB resolves, pulls, verifies, adopts and loads in the time a 12B spends on its first file.

Files

File	Purpose
`gpt2_small_fp32.safetensors`	Weights: FP32, unquantized
`gpt2_tokenizer.bin`	Mila tokenizer (BPE, 50257 entries)
`mila.json`	Manifest: file digests, quantization, minimum Mila version
`LICENSE`	OpenAI's Modified MIT License for GPT-2

Use

From the Mila chat harness — /model install works, /model load will refuse it, and that refusal is the documented behaviour above:

/model install gpt2-small

From the library:

cpp

ModelStore store;

// Pull once. This is the only thing here that touches the network.
const auto hub = makeDefaultModelHub();
ModelResolver resolver( store, *hub );
resolver.pull( "gpt2-small", std::string( kDefaultHubOwner ) );

// Load from the store thereafter -- no network, and no manifest fetch.
const auto model = store.locate( "gpt2-small" );

Installing is a deliberate step, and it is the only one that touches the network. It verifies each file against the digest in mila.json and leaves it in a content-addressed local store; every load afterwards reads the store and nothing else.

No token is required — this repository is public and ungated.

Precision

FP32, exactly as released. There is no quantized build here, and the name carries no precision suffix for that reason: the other models in this catalogue are pre-quantized deployment builds, and this one is not.

FP32 is also what Mila's training path is validated against, which is the second reason to leave it alone. A reader comparing Mila's arithmetic against a reference wants the reference unmodified.

Modifications from the base model

—Repacked from the HuggingFace checkpoint into Mila's tensor layout, then into safetensors
—Tokenizer converted to Mila's format
—No quantization, fine-tuning, distillation, or other change to what the model learned

The file opens in any safetensors reader — shapes, dtypes and metadata are all inspectable — but loading it as a model requires Mila.

License and attribution

The modification from stock MIT is a carve-out in the model's favour: the copyright and permission notices "need not be included with content created by the Software." OpenAI asks that GPT-2 be used responsibly and that generated content be clearly indicated as such.