AutoResearch-fineweb-edu-100b-shuffle-depth8

AutoResearch-fineweb-edu-100b-shuffle-depth8 is a 285.2M parameter decoder-only Transformer trained from scratch on fineweb-edu-100b-shuffle.

This model is part of the AutoResearch project, which focuses on training, evaluating, and releasing efficient language models with reproducible research workflows.

Overview

This is a 8-layer decoder-only Transformer trained on the fineweb-edu-100b-shuffle dataset for 0.2 hours of wall-clock training time. The model achieves a validation bits-per-byte (val_bpb) of 1.657887 (perplexity: 3.1555) on the held-out validation set.

References

Papers

NanoGPT / NanoChat architecture patterns

Datasets

Training: fineweb-edu-100b-shuffle
Tokenizer: climbmix-400b-shuffle

Related Projects

karpathy/autoresearch

WANDB Run

https://wandb.ai/dustinsdelivery-delivery-ai/autoresearch/runs/1btdwp5n

Highlights

Trained from scratch
285.2M parameters
Trained on 17.8M tokens (34 steps)
8-layer decoder-only Transformer with sliding window attention
RoPE positional encoding, RMSNorm, ReLU² activation
MuonAdamW optimizer (Muon for matrices, AdamW for embeddings)
Mixture of Experts (8 routed + 1 shared, top-2 routing)
Hugging Face Transformers compatible

Model Architecture

Property	Value
Architecture	Decoder-only Transformer
Parameters	285,245,968 (285.2M)
Layers	8
Hidden Size	512
Attention Heads	4
KV Heads	4
Head Dimension	128
Feed Forward Size	2048 (MoE: 8 experts, 1 shared, top-2)
Context Length	2048
Vocabulary Size	16,384
Positional Encoding	RoPE
Activation	ReLU²
Normalization	RMSNorm
Window Pattern	SSSL
Weight Tying	No

Training

This model was trained from scratch for 0.2 hours (620s) of wall-clock training time.

Training Configuration

Setting	Value
Optimizer	MuonAdamW (Muon + AdamW)
Precision	torch.bfloat16
Learning Rate	0.04 (matrix) / 0.6 (embedding)
Weight Decay	0.2
Batch Size	4 × 2048 = 8,192 tokens/step
Gradient Accumulation	64 steps
Total Batch Size	524,288 tokens
Context Length	2048
Vocabulary	16,384 tokens (BPE)
LR Scheduler	Linear warmdown (50%)
Activation Checkpointing	Enabled

Hardware

GPU: NVIDIA GeForce RTX 4060 Ti
VRAM: 16.0 GB
Peak VRAM Used: 6.3 GB
MFU: 34.96%
Framework: PyTorch 2.9.1+cu128

Dataset

Name: fineweb-edu-100b-shuffle
Language: English

Preprocessing

Data is packed into fixed-length sequences of 2048 tokens using the nanochat-compatible BPE tokenizer (16,384 vocabulary, 9 special tokens). No additional filtering or deduplication is applied beyond what is in the source dataset.

Intended Use

This model is intended for:

Educational purposes and research
Text generation experiments
Studying small language model training dynamics

Not recommended for:

Production use or safety-critical applications
Tasks requiring factual accuracy

Evaluation

Results

Metric	Score
Validation BPB	1.657887
Perplexity	3.1555
Peak VRAM	6.3 GB
MFU	34.96%

Example Generations

Example 1

Prompt

Once upon a time,

Generation

Once upon a time, set of ideas and how do it really might be.<|bos|>Effin (spot) is the most prevalent in or the first place for many years at many years, sometimes white people, are being the first white man—stant, during

Example 2

Prompt

A lonely dragon

Generation

A lonely dragon frog, who never bother him. It was a good tornado. He had to himself. To show you the woman, the women who used to use a person to accept this a man who worked hard to leave as she would come to with the

Example 3

Prompt

The opposite of boy is

Generation

The opposite of boy is in a man who is in a wooden sow. The house is right to look for the jar of fine pieces or don’t think of the car
A man is things in the history of your bag. In one another, you can probably make

Example 4

Prompt

The opposite of queen is

Generation

The opposite of queen is being for the queen, and the Kara-Fern madrebellia of his wife and his native to the Ecca.
In 191864, the Spirit of the median Paran of Franklin, was a violent,

Example 5

Prompt

My name is

Generation

My name is the first known as a Jewish scientist at the beginning of the month. The word “the earth is to be?” is still the story of the Lord, that Jesus was that when the prayer made, our heaven to the first of God

Usage

import torch
import pickle
import json
from train import GPT, GPTConfig, Tokenizer

# Load config
with open('config.json', 'r') as f:
    config_dict = json.load(f)
config = GPTConfig(**{k: v for k, v in config_dict.items() if k in GPTConfig.__dataclass_fields__})

# Load model
model = GPT(config)
state_dict = torch.load('model.pt', map_location='cpu')['state_dict']
model.load_state_dict(state_dict)
model.eval()

# Load tokenizer
with open('tokenizer.pkl', 'rb') as f:
    tokenizer = pickle.load(f)

# Generate
prompt = 'Once upon a time, '
input_ids = tokenizer.encode(prompt)
x = torch.tensor([input_ids], dtype=torch.long)
with torch.no_grad():
    for _ in range(50):
        logits = model(x)
        probs = torch.softmax(logits[:, -1, :] / 0.8, dim=-1)
        next_token = torch.multinomial(probs, num_samples=1)
        input_ids.append(next_token.item())
        x = torch.tensor([input_ids], dtype=torch.long)
print(tokenizer.decode(input_ids))

Repository Structure

model.pt                  # Model weights
config.json               # Model architecture config
dataset.txt               # Dataset name used for training
token_bytes.pt            # Token byte mappings
tokenizer.pkl             # Trained BPE tokenizer
tokenizer_config.json     # Tokenizer configuration
training_metrics.json     # Training metrics
README.md                 # This file

Limitations

Small model size limits language understanding and coherence
Trained on a single dataset (TinyStories) — limited domain
Fixed time budget training — not fully trained to convergence
No RLHF or safety alignment

Ethical Considerations

This is a research artifact, not a production model
The training data consists of synthetic stories (GPT-4 generated)
No harmful content filtering was applied
Intended for research and educational use only

Citation

@misc{autoresearch_fineweb-edu-100b-shuffle_depth8,
  title={AutoResearch-fineweb-edu-100b-shuffle-depth8},
  author={Dustin Loring},
  year={2026},
  howpublished={\url{https://huggingface.co/quik-models/colorful-shadow-83}}
}}

Version History

Version	Date	Notes
v1.0	2026-07-29	Initial release

Acknowledgements

Built with the AutoResearch training framework.

Thanks to:

Hugging Face
PyTorch
The creators of the TinyStories dataset
The open-source AI research community

License

This model is released under the MIT License unless otherwise specified.

Downloads last month: 18