SmallScale/Simple-Stories-Hindi-20M-GGUF

Simple-Stories-Hindi-22M-GGUF

This repository contains the GGUF quantization (FP16) for SmallScale/Simple-Stories-Hindi-20M, a 22M parameter decoder-only Transformer trained from scratch on Hindi simple stories.

Model Files

File Name	Quantization	Size	Description
Simple-Stories-Hindi-22M-F16.gguf	F16 (16-bit Float)	~42.8 MB	Full 16-bit float precision GGUF format

Model Architecture Details

Parameter	Value
Parameters	22.3M total (20M non-embedding)
Architecture	LLaMA-style (RoPE + SwiGLU + RMSNorm)
Context Length	512 tokens
Embedding Dim (d_model)	384
FFN Hidden Dim	1024
Layers	10
Attention Heads	8
Vocabulary	6,000 (SentencePiece Unigram)

Quick Start / Usage

1. Using with llama.cpp CLI

./llama-cli -m Simple-Stories-Hindi-22M-F16.gguf \
    -p "एक समय की बात है" \
    -n 150 \
    --temp 0.8 \
    --top-k 40 \
    --top-p 0.95

2. Using with llama-cpp-python

from llama_cpp import Llama

llm = Llama(
    model_path="Simple-Stories-Hindi-22M-F16.gguf",
    n_ctx=512,
    verbose=False
)

output = llm(
    "एक समय की बात है",
    max_tokens=150,
    temperature=0.8,
    top_k=40,
    top_p=0.95
)

print(output["choices"][0]["text"])

3. Using in Browser (WebAssembly / Wllama)

Try the live browser demo powered by WebAssembly & WebGPU: SmallScale/simple-stories-20m-webgpu

License

MIT