README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language: en
datasets:
  - roneneldan/TinyStories
tags:
  - tiny
  - slm
  - small-language-model
  - sub-1m
  - from-scratch
  - gqa
  - llama
metrics:
  - perplexity
---

# TinyStories-40M

A **39.6M-parameter** LLaMA-style transformer trained **from scratch** on the
full TinyStories corpus. Published as a from-scratch training demonstration at
the ~40M scale — it is *not* a coherent story generator (see Quality below).

## Architecture

| Parameter | Value |
|---|---|
| Layers | 12 |
| Hidden size | 512 |
| Attention heads (Q) | 8 |
| Attention heads (KV) | 4 (GQA 2:1) |
| Head dim | 64 |
| FFN (SwiGLU) | 1408 |
| Vocab size | 8192 (BPE) |
| Context length | 512 |
| RoPE θ | 10000 |
| Norm | RMSNorm (pre-norm) |
| Tied embeddings | Yes |
| Precision | FP32 |

**Total parameters: 39,596,544** (86 tensors, head weight tied to the token
embedding). Verified against the published `model.safetensors`.

## Training

- **Data:** roneneldan/TinyStories (~1.9B chars, ~490M tokens at seq 512)
- **Steps:** 15,000
- **Batch size:** 64 sequences × 512 tokens (32,704 tok/step)
- **Optimizer:** AdamW, lr 3e-4, cosine decay (min-lr-frac 0.1), 300-step warmup
- **Hardware:** RTX 5090 (32 GB), ~4 hours
- **Final train loss:** 3.23 (step 15000)

## Quality

Honest picture, from running the published weights (temp 0.7, top-k 50):

- On the canonical prompt **"Once upon a time"** the model produces a
  story-like first sentence, then degrades:

  > "Once upon a time, his mom and his clapped and cheered for him. They all
  > enjoyed spending the rest of their special day in the park, his mom's light
  > and a Stop being himself."

- On other prompts it is weaker still — often a single short phrase or an
  immediate stop:

  > "In the forest" → "In the forest things."
  > "She opened the door" → "She opened the door."

- Longer generations drift into incoherent, non-grammatical text.

This is what a 40M from-scratch model on a single 490M-token corpus
demonstrably does: it learns the surface distribution of story text (word
order, names, punctuation, the "Once upon a time" register) but does not
sustain coherent narrative. It is published as a training demonstration, not
as a usable story generator.

## Usage

This is a custom `transformers` model — load with `trust_remote_code=True`:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Compactbot/tinystories-40m", trust_remote_code=True, torch_dtype=torch.float32
)
tok = AutoTokenizer.from_pretrained("Compactbot/tinystories-40m")

prompt = "Once upon a time"
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=100, temperature=0.7, top_k=50)
print(tok.decode(out[0], skip_special_tokens=True))
```

Note: the bundled `generate()` always samples (no `do_sample` flag); pass
`temperature` and `top_k` to control it.

## Notes

- From-scratch training: initialized randomly, trained end-to-end on
  TinyStories. No pre-training from any other model.
- GQA (Grouped Query Attention) 2:1, SwiGLU, RoPE — LLaMA-2 recipe at 40M scale.
- ~151 MB in FP32 — under 200 MB.

## Files

| file | what |
|------|------|
| `model.safetensors` | 39,596,544 params, 86 tensors, FP32 |
| `config.json` | architecture + training metadata |
| `modeling_tinystories.py` | the model class (loaded via `trust_remote_code`) |
| `tokenizer.json` | BPE-8k tokenizer (HF `tokenizers` format) |