persian-lm-from-scratch

Persian GPT - 110M parameters, trained from scratch

A decoder-only transformer language model trained entirely from scratch on Persian Wikipedia text. No pretrained weights, no pretrained tokenizer, no distillation — the model, the tokenizer, and the training corpus were all built for this project.

Model summary

Evaluation

11.8x improvement over the 5-gram baseline. Best val loss: 2.6391. Test loss: 2.6756.

Training details

• Corpus: Persian Wikipedia (cleaned) — 855,724 documents, 244M tokens
• Tokenizer: 32,000-token BPE trained from scratch, ZWNJ preserved
• Hardware: Google Colab Pro, NVIDIA A100 80GB
• Optimizer: AdamW (beta1=0.9, beta2=0.95, weight decay 0.1)
• LR schedule: Linear warmup (500 steps) + cosine decay to 10% of peak
• Peak LR: 3e-4
• Effective batch: 64 (32 x 2 accumulation)
• Epochs: 2 (7,450 optimizer steps)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "aradmanamnaoon/persian-lm-from-scratch",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "aradmanamnaoon/persian-lm-from-scratch",
    trust_remote_code=True,
)

Limitations

• Register: trained only on Wikipedia — produces encyclopedic prose, not conversational Persian.
• Factual accuracy: no grounding mechanism. Do not use for factual QA.
• Scale: 244M tokens for 110M params. Larger models will outperform this one.
• License: inherits CC-BY-SA-3.0 from the training data.

The safetensors file is updated and it is a slightly better version

Citation

@misc{persian_gpt_110m, title = {Persian GPT - 110M parameters, trained from scratch}, author = {Arad Manamnaoon}, year = {2026}, url = {https://huggingface.co/aradmanamnaoon/persian-lm-from-scratch} }