Instructions to use Efe-Tatar/base-100m-fineweb-edu with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Efe-Tatar/base-100m-fineweb-edu with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Efe-Tatar/base-100m-fineweb-edu")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Efe-Tatar/base-100m-fineweb-edu")
model = AutoModelForCausalLM.from_pretrained("Efe-Tatar/base-100m-fineweb-edu", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Efe-Tatar/base-100m-fineweb-edu with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Efe-Tatar/base-100m-fineweb-edu"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Efe-Tatar/base-100m-fineweb-edu",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Efe-Tatar/base-100m-fineweb-edu

SGLang

How to use Efe-Tatar/base-100m-fineweb-edu with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Efe-Tatar/base-100m-fineweb-edu" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Efe-Tatar/base-100m-fineweb-edu",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Efe-Tatar/base-100m-fineweb-edu" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Efe-Tatar/base-100m-fineweb-edu",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Efe-Tatar/base-100m-fineweb-edu with Docker Model Runner:
```
docker model run hf.co/Efe-Tatar/base-100m-fineweb-edu
```

base_100m: a Llama-style model pretrained from scratch

A 100M-parameter decoder-only language model whose architecture, tokenizer, data pipeline and training loop were written from scratch in PyTorch, then exported to the standard LlamaForCausalLM format. Code, tests and training logs: https://github.com/efe-tatar/transformer-pytorch

This is a base model: it continues text, it does not follow instructions or chat.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tok = AutoTokenizer.from_pretrained("Efe-Tatar/base-100m-fineweb-edu")
model = AutoModelForCausalLM.from_pretrained("Efe-Tatar/base-100m-fineweb-edu")

# documents started with <|endoftext|> during training, so prompts should too
ids = tok("<|endoftext|>Photosynthesis is the process by which", return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=60, do_sample=True, temperature=0.8, top_k=50)
print(tok.decode(out[0][1:]))

Model


Parameters	100.7M (25.2M of them the tied input/output embedding)
Shape	768 hidden, 12 layers, 12 attention heads, 4 KV heads (grouped-query attention), 1024 context
Blocks	pre-norm RMSNorm, RoPE (base 10000), SwiGLU feed-forward (2048 hidden)
Tokenizer	byte-level BPE, 32,768 tokens, trained from scratch on 300M characters of FineWeb-Edu

The export is checked against the original implementation: identical token ids, logits equal to within 1e-5, identical greedy generation.

Training


Data	2.99B tokens of FineWeb-Edu (`sample-10BT`), one epoch
Optimizer	AdamW (β = 0.9, 0.95, weight decay 0.1 on matmul weights only), gradient clipping 1.0
Schedule	0.5M tokens per step, 5700 steps, LR 6e-4, 700-step warmup, cosine decay to 6e-5
Hardware	1× A100 80GB, bf16, 4.5 hours, 184k tokens/s (42% MFU)
Final validation loss	3.223

Evaluation

HellaSwag validation (10,042 questions), scored with identical code for both models:

Model	Training tokens	acc	acc_norm
random guessing		25.0%	25.0%
base_100m	3B	27.7%	27.7%
GPT-2 124M	far more (undisclosed)	28.7%	29.6%

Limitations

Small and trained on 3B tokens: fluent and on topic, but makes up facts and repeats itself.
English only, educational web text only (FineWeb-Edu); weak on code.
No safety tuning or instruction tuning of any kind.

Downloads last month: 44

Safetensors

Model size

0.1B params

Tensor type

F32

Efe-Tatar
/

base-100m-fineweb-edu