README.md

metadata

language:
  - en
license: mit
tags:
  - transformer
  - language-model
  - pytorch
  - from-scratch
  - llama
  - decoder-only
  - causal-lm
library_name: pytorch
pipeline_tag: text-generation
datasets:
  - HuggingFaceFW/fineweb-edu
model-index:
  - name: llama-124m-fineweb
    results:
      - task:
          type: text-generation
          name: Language Modeling
        dataset:
          name: FineWeb-Edu (sample-10BT)
          type: HuggingFaceFW/fineweb-edu
          config: sample-10BT
        metrics:
          - name: Validation Loss
            type: loss
            value: 3.1626
          - name: Perplexity
            type: perplexity
            value: 23.62

LLaMA-124M

A 124M parameter LLaMA-style decoder-only transformer trained from scratch on FineWeb-Edu (sample-10BT, ~9.65 billion tokens).

Every component — attention, feed-forward layers, normalization, positional encoding, training loop — is implemented from scratch in pure PyTorch. No HuggingFace transformers library.

Source code: github.com/alexgarabt/transformer

Model Architecture

Parameter	Value
Parameters	124,472,064
d_model	768
n_heads	12
n_layers	12
d_ff	2,560
max_seq_len	1,024
Activation	SwiGLU
Normalization	RMSNorm (pre-norm)
Position encoding	Learned
Weight tying	Yes (embedding ↔ LM head)
Vocab size	32,000 (SentencePiece BPE)
Dropout	0.0
Bias	No

Training Details

Detail	Value
Dataset	FineWeb-Edu sample-10BT (~9.65B tokens)
Epochs	1
Effective batch size	128 × 1,024 × 4 = 524,288 tokens/step
Optimizer	AdamW (lr=6e-4, betas=(0.9, 0.95), weight_decay=0.1)
LR schedule	Cosine decay with 750 warmup steps
Precision	bfloat16 (mixed precision)
Attention	Flash Attention (F.scaled_dot_product_attention)
Gradient clipping	1.0 (max norm)
Hardware	1× NVIDIA H200 (~110 GB VRAM)
Training time	~6 hours
Estimated cost	~$24
Final val loss	3.1626
Final perplexity	~23.6
Seed	42

Usage

Quick Start

git clone https://github.com/alexgarabt/transformer
cd transformer
uv sync --extra data
uv run python scripts/inference.py --prompt "The meaning of life" --device cpu

Download Weights Manually

from huggingface_hub import hf_hub_download

model_path = hf_hub_download("alexgarabt/llama-124m-fineweb", "model.pt")
params_path = hf_hub_download("alexgarabt/llama-124m-fineweb", "params.json")
tokenizer_path = hf_hub_download("alexgarabt/llama-124m-fineweb", "tokenizer.model")

Load and Generate (Python)

import torch
from transformer.config import TransformerLMConfig
from transformer.models import TransformerLM
from transformer.data import Tokenizer
from transformer.generation import generate

# Load config
import json
with open(params_path) as f:
    params = json.load(f)

config = TransformerLMConfig(**params["model"])
model = TransformerLM(config)

# Load weights
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Load tokenizer and generate
tokenizer = Tokenizer(tokenizer_path)
prompt_ids = torch.tensor([tokenizer.encode("Once upon a time", add_bos=True)])
output = generate(model, prompt_ids, max_new_tokens=100, temperature=0.8, top_p=0.95)
print(tokenizer.decode(output[0].tolist()))

Files in This Repository

File	Description	Size
model.pt	Model weights (clean state dict)	~500 MB
params.json	Model & training configuration	~1 KB
tokenizer.model	SentencePiece BPE tokenizer	~786 KB
tokenizer.vocab	Vocabulary file	~503 KB
config/llama_124M.json	Training configuration	~1 KB
runs/	TensorBoard training logs	~95 MB

Limitations

• Base model only — no instruction tuning (SFT), no RLHF, no DPO. The model generates coherent English text but does not follow instructions or answer questions reliably.
• Small model — 124M parameters is relatively small. Generation quality is limited compared to larger models.
• Single epoch — trained for only one pass over the dataset.
• English only — the tokenizer and training data are English.
• Max 1024 tokens — the model's positional encoding supports sequences up to 1024 tokens.