poem-llm-small

🌹 Poetry SLM — Small Language Model

A decoder-only Small Language Model (SLM) inspired by GPT-2 architecture, trained from scratch on poetry.

Model Details

Training Data

• Kaggle Poems Dataset (~15,973 poems)
• Shakespeare Sonnets (Project Gutenberg)
• Shakespeare Poetry (Hugging Face merve/poetry)

How to Use

Step 1 — Install dependencies

pip install torch sentencepiece huggingface_hub

Step 2 — Inference code

import torch
import torch.nn as nn
from torch.nn import functional as F
import sentencepiece as spm
from huggingface_hub import hf_hub_download
import re, io

# ============================================
# CONFIG
# ============================================
REPO_ID = "Jyotiprakash4357/poem-llm-small"
embed_size = 256
n_layers = 6
n_heads = 6
context = 256
BIAS = True
dropout = 0.0
device = "cuda" if torch.cuda.is_available() else "cpu"

# ============================================
# LOAD TOKENIZER
# ============================================
tokenizer_path = hf_hub_download(
    repo_id=REPO_ID,
    filename="poems_tokenizer.model"
)
sp = spm.SentencePieceProcessor(model_file=tokenizer_path)
vocab_size = sp.get_piece_size()
encode = lambda s: sp.Encode(s)
decode = lambda l: sp.Decode(l)

# ============================================
# MODEL ARCHITECTURE
# ============================================
class Head(nn.Module):
    def __init__(self, head_size):
        super().__init__()
        self.queries = nn.Linear(embed_size, head_size, bias=BIAS)
        self.keys = nn.Linear(embed_size, head_size, bias=BIAS)
        self.values = nn.Linear(embed_size, head_size, bias=BIAS)
        self.register_buffer('tril', torch.tril(torch.ones(context, context)))
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        BS, SL, VS = x.shape
        q = self.queries(x)
        k = self.keys(x)
        v = self.values(x)
        attn_w = q @ k.transpose(-2, -1) * k.shape[-1]**-0.5
        attn_w = attn_w.masked_fill(self.tril[:SL, :SL]==0, float('-inf'))
        attn_w = F.softmax(attn_w, dim=-1)
        attn_w = self.dropout(attn_w)
        return attn_w @ v

class Multihead(nn.Module):
    def __init__(self, n_heads, head_size):
        super().__init__()
        self.heads = nn.ModuleList([Head(head_size) for _ in range(n_heads)])
        self.combine = nn.Linear(head_size * n_heads, embed_size, bias=BIAS)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = torch.cat([head(x) for head in self.heads], dim=-1)
        x = self.combine(x)
        return self.dropout(x)

class ForwardLayer(nn.Module):
    def __init__(self, embed_size):
        super().__init__()
        self.network = nn.Sequential(
            nn.Linear(embed_size, 6*embed_size, bias=BIAS),
            nn.GELU(),
            nn.Linear(6*embed_size, embed_size, bias=BIAS),
            nn.Dropout(dropout)
        )
    def forward(self, x):
        return self.network(x)

class Block(nn.Module):
    def __init__(self, n_heads):
        super().__init__()
        head_size = embed_size // n_heads
        self.ma = Multihead(n_heads, head_size)
        self.feed_forward = ForwardLayer(embed_size)
        self.ln1 = nn.LayerNorm(embed_size)
        self.ln2 = nn.LayerNorm(embed_size)

    def forward(self, x):
        x = x + self.ma(self.ln1(x))
        x = x + self.feed_forward(self.ln2(x))
        return x

class GPT(nn.Module):
    def __init__(self):
        super().__init__()
        self.embeddings = nn.Embedding(vocab_size, embed_size)
        self.positions = nn.Embedding(context, embed_size)
        self.blocks = nn.Sequential(*[Block(n_heads) for _ in range(n_layers)])
        self.ln = nn.LayerNorm(embed_size)
        self.final_linear = nn.Linear(embed_size, vocab_size, bias=BIAS)
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, input, targets=None):
        loss = None
        BS, SL = input.shape
        emb = self.embeddings(input)
        pos = self.positions(torch.arange(SL, device=device))
        x = emb + pos
        x = self.blocks(x)
        x = self.ln(x)
        logits = self.final_linear(x)
        if targets is not None:
            BS, SL, VS = logits.shape
            logits = logits.view(BS*SL, VS)
            targets = targets.view(BS*SL)
            loss = F.cross_entropy(logits, targets)
        return logits, loss

    def generate(self, input, max=200, temperature=0.8, 
                 top_k=40, repetition_penalty=1.3):
        for _ in range(max):
            input_crop = input[:, -context:]
            logits, _ = self(input_crop)
            logits = logits[:, -1, :]
            for token_id in set(input[0].tolist()):
                logits[0, token_id] /= repetition_penalty
            logits = logits / temperature
            top_k_logits, top_k_indices = torch.topk(logits, top_k)
            probs = F.softmax(top_k_logits, dim=-1)
            next_token_pos = torch.multinomial(probs, num_samples=1)
            next_token = top_k_indices.gather(-1, next_token_pos)
            input = torch.cat((input, next_token), dim=1)
        return input

# ============================================
# LOAD MODEL
# ============================================
model = GPT().to(device)

model_path = hf_hub_download(
    repo_id=REPO_ID,
    filename="poems_latest.pt"
)

with open(model_path, 'rb') as f:
    buffer = io.BytesIO(f.read())

checkpoint = torch.load(buffer, map_location=device)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# ============================================
# GENERATE
# ============================================
def clean_output(text, prompt):
    text = text[len(prompt):]
    text = re.sub(r'©.*?\n', '', text)
    lines = text.split('\n')
    clean_lines = [
        l for l in lines
        if len(re.findall(r'[^\x00-\x7F]', l)) <= 2
    ]
    return '\n'.join(clean_lines).strip()

@torch.no_grad()
def generate_poem(prompt, max_tokens=200, temperature=0.8,
                  top_k=40, repetition_penalty=1.3):
    t1 = torch.tensor(encode(prompt), dtype=torch.long, device=device)
    t1 = t1[None, :]
    output = model.generate(
        t1, max=max_tokens,
        temperature=temperature,
        top_k=top_k,
        repetition_penalty=repetition_penalty
    )[0].tolist()
    return clean_output(decode(output), prompt)

# Test
print(generate_poem("The roses bloom in darkness"))
print(generate_poem("Death comes softly"))
print(generate_poem("Love is a shadow"))

Example Outputs

Prompt: "Death comes softly"

Death comes softly,
From the dark and the dim cloudless sea,
There is one thing in love, which has a soul
That once came into the world...

Limitations

• 17M training tokens is small for this architecture
• Occasionally generates incoherent text
• May repeat phrases without repetition penalty
• Not an instruction tuned model — text generation only

Try it Live

🔗 Hugging Face Space