kamal-018/Maithili_Poem_Generator

MaithiliNano — Maithili Poem Generator (Variant A)

A decoder-only Small Language Model (SLM) that generates Maithili poetry. It was pre-trained from scratch on a 45.86M-token Maithili corpus, then fine-tuned on curated Maithili poems. Part of the MaiGen project.

Model

• —Architecture: Decoder-only Transformer (RoPE, RMSNorm, SwiGLU, tied embeddings)
• —Parameters: ~29.3M
• —Layers / Heads / Dim: 8 / 8 / 512
• —Vocab / Context: 8,000 Byte-Level BPE tokens / 512
• —Checkpoint: epoch 10, step 1200, val loss ~0.70

Files

• —model.py — model class + from_pretrained()
• —model.safetensors + config.json — weights and config (auto-downloaded)
• —vocab.json + merges.txt — ByteLevelBPE tokenizer (auto-downloaded)
• —poem_finetune.pt — original PyTorch checkpoint (model_state, epoch, step, val_loss)

Quickstart

pip install torch tokenizers safetensors huggingface_hub

Put model.py in your working folder, then run:

import torch
from model import MaithiliNano

# Downloads weights, config, vocab + merges straight from HF
model = MaithiliNano.from_pretrained("kamal-018/Maithili_Poem_Generator")
model.eval()

# Tokenizer is auto-downloaded into the same cache folder
tok = model.tokenizer()


def generate(prompt, max_len=150, temperature=0.6, top_k=20):
    ids = [tok.token_to_id("<s>")] + tok.encode(prompt).ids
    idx = torch.tensor([ids])
    out = model.generate(idx, max_new_tokens=max_len,
                         temperature=temperature, top_k=top_k)
    return tok.decode(out[0].tolist())


print(generate("प्रेम"))

-This outputs a Maithili poem about love (प्रेम).\ प्रेम आ भाव प्रेम भाव ऊँचे,\ मानव के सजय छी हम संग। \ जंगली घास के मैदावाला दिस बढ़लै \ विचार हमर त्राण परिचालिनी, \ लावारपाटी में हम मतिमे सहायता \ अयली माता काली माय सेहो श्रेष्ठ \ हम लोकक सासुर फुजल छै रहब

Notes

• —The prompt is auto-prepended with <s> (BOS). Generation uses temperature 0.6, top-k 20.
• —Uses a custom from_pretrained (this repo's model.py) — it is not a transformers.AutoModel. You do not need transformers installed.
• —Full model definition and training code are in the project repo (Maithili_Poem_Generator).