A tiny (~14M parameter) GPT-2-architecture chat model, trained completely from scratch — no pretrained base model. Custom BPE tokenizer trained from zero, custom transformer trained from random initialization, then converted into a standard GPT2LMHeadModel for compatibility with the wider ecosystem.

Looking for a .gguf build? See mondk/GGUF.msh-tiny.

Limitations

Trained from random initialization on a modest amount of data with limited compute — a small educational project, not a production-quality assistant. Expect reliable chat formatting but limited/inconsistent knowledge and occasional incoherent answers.

Prompt format

<|user|>
{your message}
<|assistant|>

The model stops generating at <|end|>.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("mondk/Msh-Tiny-14M")
tokenizer = AutoTokenizer.from_pretrained("mondk/Msh-Tiny-14M")

prompt = "<|user|>\nhi\n<|assistant|>\n"
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
output = model.generate(input_ids, max_new_tokens=100, do_sample=True, temperature=0.7, top_k=40)
print(tokenizer.decode(output[0][input_ids.shape[1]:], skip_special_tokens=True))

Training data

Combining 3 well-known open instruction/chat datasets plus a small hand-written set of everyday chit-chat (greetings, thanks, small talk):

tatsu-lab/alpaca
teknium/OpenHermes-2.5
HuggingFaceH4/no_robots

mondk/Safetensors.msh-tiny

msh-tiny

Limitations

Prompt format

Usage

Training data

mondk/
Safetensors.msh-tiny