Instructions to use avneeshjadhav04/llm-from-scratch with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use avneeshjadhav04/llm-from-scratch with Transformers:

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

pipe = pipeline("text-generation", model="avneeshjadhav04/llm-from-scratch")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("avneeshjadhav04/llm-from-scratch", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use avneeshjadhav04/llm-from-scratch with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "avneeshjadhav04/llm-from-scratch"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "avneeshjadhav04/llm-from-scratch",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/avneeshjadhav04/llm-from-scratch

SGLang

How to use avneeshjadhav04/llm-from-scratch 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 "avneeshjadhav04/llm-from-scratch" \
    --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": "avneeshjadhav04/llm-from-scratch",
		"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 "avneeshjadhav04/llm-from-scratch" \
        --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": "avneeshjadhav04/llm-from-scratch",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use avneeshjadhav04/llm-from-scratch with Docker Model Runner:
```
docker model run hf.co/avneeshjadhav04/llm-from-scratch
```

LLM from Scratch (124M)

A clean, from-scratch implementation of a 124M-parameter decoder-only Transformer in PyTorch. No nn.Transformer, no shortcuts — every layer (attention, FFN, LayerNorm, embeddings) is built manually. Trained on FineWeb-Edu with mixed precision, gradient accumulation, and automatic checkpoint resume.

🚀 Live Demo: https://avneeshjadhav04--llm-api.modal.run

📂 Training Code: github.com/avneeshjadhav04/llm-from-scratch

Model Description

This is a GPT-2-style causal language model trained entirely from random initialization on the FineWeb-Edu dataset. It was built as an educational and portfolio project to demonstrate deep understanding of Transformer internals and large-scale training loops.

Property	Value
Architecture	Decoder-only Transformer (GPT-2 style)
Parameters	~124.44M
Vocab Size	50,257 (GPT-2 / tiktoken)
Context Length	1024 tokens
Hidden Size	768
Layers	12
Attention Heads	12
FFN Dimension	3,072
Activation	GELU
Normalization	Pre-LayerNorm
Weight Tying	Input / Output embeddings shared

Training Results

Metric	Value
Final validation loss	2.6943
Final validation perplexity	14.8
Training tokens	~2B (FineWeb-Edu)
Train/val split	95% / 5%
Hardware	NVIDIA A100-SXM4-80GB
Wall time	~5 hours
Framework	PyTorch 2.10 + torch.compile
Precision	FP16 mixed precision

Intended Use

Educational: Understanding how GPT-2-scale models work under the hood.
Research: Baseline for data-efficiency or architectural ablation studies.
Text Generation: Short-form English text completion with temperature / top-k / top-p sampling.

Limitations

Scale: 124M params is small by modern standards. Output can be repetitive on long generations.
Dataset: Trained exclusively on FineWeb-Edu (filtered web text). May reflect web biases.
Language: English only.
Length: Best results under 50–100 generated tokens. Longer generations may loop.
Optimization: Trained to Chinchilla-optimal scale (~20× parameters in tokens).

Architecture Highlights

Manual Transformer: Every component coded from scratch (no nn.Transformer).
Pre-Norm + Weight Tying: Stable training, fewer parameters, better perplexity.
Mixed Precision: FP16 via torch.amp for ~1.5× speedup.
Session-Safe: Auto-resumes from any checkpoint (safe for Colab / Kaggle timeouts).
Custom Sampling: Temperature, top-k, and top-p nucleus sampling implemented manually.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("your-username/llm-from-scratch-124m")
tokenizer = AutoTokenizer.from_pretrained("your-username/llm-from-scratch-124m")
inputs = tokenizer("The future of artificial intelligence is", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.8, top_k=40)
print(tokenizer.decode(outputs[0]))
Citation
If you use this model or code, please cite:
@misc{llm-from-scratch-2024,
  author = {Avneesh Jadhav},
  title = {LLM from Scratch: A 124M-Parameter Transformer in PyTorch},
  year = {2024},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/avneeshjadhav04/llm-from-scratch}}
}

Acknowledgments

Inspired by Andrej Karpathy's nanoGPT (https://github.com/karpathy/nanoGPT)
Transformer architecture from Attention Is All You Need (https://arxiv.org/abs/1706.03762)
FineWeb-Edu dataset by HuggingFace (https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)

License

Code: MIT License
Dataset: FineWeb-Edu is released under ODC-By v1.0 (https://opendatacommons.org/licenses/by/1-0/)

Downloads last month: -; Downloads are not tracked for this model. How to track

Dataset used to train avneeshjadhav04/llm-from-scratch

Space using avneeshjadhav04/llm-from-scratch 1

Paper for avneeshjadhav04/llm-from-scratch

Attention Is All You Need

Paper • 1706.03762 • Published Jun 12, 2017 • 122