---
license: apache-2.0
language:
- en
tags:
- text-generation
- gpt
- nanogpt
- pytorch
- from-scratch
- small-language-model
- slm
- pretraining
- muon-optimizer
- single-gpu
- base-model
- research
pipeline_tag: text-generation
---

![image](https://cdn-uploads.huggingface.co/production/uploads/6a487897487a4a16ab1124f2/Wubj03NXiHQmgfrGzExNZ.png)

# GTM-v2-base by OpenGCM

A ~120M parameter, decoder-only GPT-style **base** language model, trained
from scratch on a single RTX Pro 6000, on roughly 4 billion tokens streamed
from a mix of open web/educational/math corpora.

**This is a base (pretrained) model, not an instruction-tuned or chat
model.** It completes/continues text; it does not reliably follow
instructions or answer questions directly. It also has **no fine-tuning
for factual accuracy** — it can and does confidently generate fluent,
plausible-sounding but factually incorrect content. Treat it as a small,
from-scratch research/hobby model, not a production system.

A separate instruction-tuned variant (SFT on UltraChat) is planned as a
follow-up release; this repo is the base-only checkpoint.

## Model details

- **Architecture**: nanoGPT-style decoder-only transformer
  - 14 layers, 8 attention heads, 704 embedding dim
  - ~119.4M parameters (weight-tied embeddings/output head)
  - Context length: 1024 tokens
  - Uses PyTorch's fused `scaled_dot_product_attention` (flash-attention kernel)
- **Tokenizer**: `tiktoken` GPT-2 BPE encoding (`tiktoken.get_encoding("gpt2")`),
  vocab size 50,257. No custom tokenizer was trained.
- **Optimizer**: Muon (for 2D weight matrices) + AdamW (for embeddings,
  layernorms, biases) — a hybrid setup, following the approach popularized
  in recent efficient-pretraining recipes.
- **Precision**: trained with bf16 autocast; released weights are fp32.
- **Training tokens seen**: ~4.16B tokens (40,690 steps x effective batch
  100 x 1024 context)
- **Training data**: streamed via HuggingFace `datasets`, mixed and
  tokenized on the fly (no fixed local copy of the source datasets):
  - FineWeb-Edu (`HuggingFaceFW/fineweb-edu`, `sample-10BT`) — 45%
  - Cosmopedia-v2 (`HuggingFaceTB/cosmopedia-v2`) — 30%
  - FineMath (`HuggingFaceTB/finemath`, `finemath-4plus`) — 15%
  - FineWeb (`HuggingFaceFW/fineweb`, `sample-10BT`) — 10%

  Note: code data (The Stack v2, StarCoderData, the-stack-smol) was
  intentionally left out of this mix — every BigCode-hosted code corpus we
  checked is gated behind a terms-of-use click-through on HuggingFace, so
  none of them are pulled in without the user separately accepting those
  terms and authenticating.

This is the second model in this series. An earlier ~100M parameter
version (GTM-v1-base, trained on ~3B tokens with the same data mix and
optimizer) informed several of this model's design choices, including the
Muon/AdamW optimizer split and the corpus weighting above.

## Benchmarks

Evaluated via likelihood-scoring (comparing the model's per-token loss
across candidate answers, no generation/sampling involved), 200 examples
per benchmark, final checkpoint at step 40,690 (~4.16B tokens seen):

| Benchmark      | GTM-v2-base | GTM-v1-base | GPT-2 (124M) | Random baseline |
|----------------|:-----------:|:-----------:|:------------:|:----------------:|
| HellaSwag      | 33.0%       | 32.0%       | ~28-29%      | 25%              |
| ARC-Easy       | 43.0%       | 42.5%       | ~39.2%       | ~25%             |
| ARC-Challenge  | 24.0%       | 26.0%       | ~22.5%       | ~25%             |

Corpus perplexity (per-source, lower is better) on freshly streamed,
non-training-shard text:

| Source         | Perplexity |
|----------------|:----------:|
| Cosmopedia-v2  | 11.80      |
| FineMath       | 13.04      |
| FineWeb-Edu    | 27.68      |
| FineWeb        | 47.34      |
| **Overall**    | **21.19** (vs. GTM-v1-base's 22.87) |

GTM-v2-base improves on GTM-v1-base (an earlier ~100M param, ~3B token
version of this same series) on HellaSwag, ARC-Easy, and overall
perplexity, but is *not* a clean sweep — it's slightly behind v1 on
ARC-Challenge (24.0% vs. 26.0%), though both remain above GPT-2's ~22.5%
reference. ARC-Challenge is a small (200-example), noisy evaluation at
this scale, so a few-point gap in either direction shouldn't be read as a
strong signal on its own.

GPT-2 (124M) reference numbers are from a from-scratch reproduction
verified to match official GPT-2 evaluation results. GTM-v2-base was
trained on roughly 1/10th of GPT-2's training tokens (~4B vs. ~40B), on a
single consumer/workstation GPU rather than a large multi-GPU cluster.

**These numbers should be read as "beats a 2019 baseline on a few narrow
multiple-choice benchmarks," not "as good as or better than GPT-2 in
general."** GPT-2 was trained and evaluated far more broadly; free-form
generation quality, factual grounding, and behavior outside these specific
benchmark formats have not been rigorously compared.

## Known limitations

- **No reliable factual recall.** E.g. prompted with "The capital of France
  is", the model does not consistently produce "Paris" — it may produce
  fluent but factually invented content instead. This is expected: the
  training corpus is not dense in discrete factual content, and this
  model's parameter/token budget is small for reliably memorizing specific
  facts.
- **No code capability.** No code-specific training data was included (see
  above). The model can produce code-*shaped* text (recognizing "write
  code" prompts and using plausible syntax) but not functionally correct
  code.
- **Not instruction-tuned.** Does not follow instructions or answer
  questions in a chat-like way — it continues text. (A separate SFT/chat
  release is planned; this is the base-only checkpoint.)
- **Repetition tendency.** Base generation (greedy or low-temperature
  sampling) can fall into repetition loops or lock onto structural
  templates (e.g. numbered-list/fill-in-the-blank formatting) when it
  lacks a confident continuation. A repetition penalty (see usage example)
  substantially reduces this.

## Usage

Requires `model.py` (included in this repo) alongside the checkpoint —
this is a plain PyTorch model, not a `transformers` `AutoModel`.

```bash
pip install torch safetensors tiktoken
```

```python
import json
import torch
from safetensors.torch import load_file
from model import GPT, GPTConfig

with open("config.json") as f:
    cfg_dict = json.load(f)
config = GPTConfig(
    vocab_size=cfg_dict["vocab_size"], block_size=cfg_dict["block_size"],
    n_layer=cfg_dict["n_layer"], n_head=cfg_dict["n_head"],
    n_embd=cfg_dict["n_embd"], dropout=cfg_dict["dropout"], bias=cfg_dict["bias"],
)
model = GPT(config)
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()

import tiktoken
enc = tiktoken.get_encoding("gpt2")

prompt = "Once upon a time,"
ids = enc.encode_ordinary(prompt)
x = torch.tensor([ids], dtype=torch.long)

with torch.no_grad():
    out = model.generate(
        x, max_new_tokens=128, temperature=0.8, top_k=50,
        eot_token=enc.eot_token, repetition_penalty=1.3,
    )

print(enc.decode(out[0].tolist()))
```

## License

Apache 2.0 for this repo's contents (model weights, `model.py`, and this
README). The underlying training data retains its own licenses regardless
of the license on this trained model — FineWeb, FineWeb-Edu, Cosmopedia-v2,
and FineMath are each ODC-BY-1.0 (see their respective HF dataset cards for
full terms and attribution requirements). This repo distributes model
weights, not the training data itself.