MTLM2-40M is a highly experimental, tiny language model (~40 Million parameters) designed as a research artifact to explore the lower bounds of language modeling capabilities.

The primary research question driving this model was:

Can a microscopic model (40M params), when saturated with a large amount of data (14B tokens), generate minimally coherent narrative text?

The answer is yes. Validated by a Perplexity of 54.21 on WikiText-2, the model demonstrates surprising structural and narrative cohesion for its size.

📊 Model Details

Architecture: TinyGPT (Custom Llama-style architecture with tweaks for small-scale efficiency).
Parameters: ~41.5 Million.
Training Data: 14 Billion Tokens (English).
Training Framework: JAX / Flax (XLA).
Hardware: Google Cloud TPU v5e-8.
Training Time: ~6 Hours.
Precision: FP32 (Exported to Safetensors).

📉 Benchmarks

Benchmark	Metric	Result
WikiText-2	Perplexity (PPL)	54.21

Note: Evaluation performed using sliding window approach. The low PPL confirms strong grammatical alignment despite the small parameter count.

🧪 Capabilities & Limitations

What it CAN do:

Narrative Cohesion: It can generate grammatically correct English sentences that flow logically.
High Saturation: Trained on ~350 tokens per parameter (far exceeding Chinchilla scaling laws), making it extremely "dense".

What it CANNOT do:

Reasoning/Logic: It cannot solve math problems, code, or answer complex logical queries.
Instruction Following: It has NOT undergone Supervised Fine-Tuning (SFT). It is a base completion model.

Usage

This model requires trust_remote_code=True.

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Madras1/MTLM2-40M"

model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")

prompt = "The future of AI is"
inputs = tokenizer(prompt, return_tensors="pt")

output = model.generate(**inputs, max_new_tokens=50, do_sample=True, temperature=0.7)
print(tokenizer.decode(output[0]))

Author Developed by Madras1 (Gabriel).