You can deploy this model and start earning money today!

Kush26/
rhapsody-84m-haiku-poet

Name: rhapsody-84m-haiku-poet
Availability: InStock

text-generation

taucris/haiku_333K

Kush26/rhapsody-84m-pretrain

finetune

mit

Views

No views yet

Rhapsody Constraint Poet (Haiku Generator)

Rhapsody Constraint Poet is an 84-million parameter Transformer language model fine-tuned to generate structured, topic-constrained 5-7-5 haikus.

It is fine-tuned from the base Rhapsody TextLM (84M) on a subset of the taucris/haiku_333K dataset.

Model Details

Developer: Kush Singh
Model Type: Decoder-only Transformer
Fine-tuned From: Rhapsody TextLM (84M base)
Language: English
License: MIT
Task: Topic-constrained 5-7-5 Haiku Generation
Inference Mode: Autoregressive Generation (temperature = 0.7, repetition_penalty = 1.15)

Fine-tuning Details

The model was fine-tuned for sequence-to-sequence style prompt-response generation:

Training Format: Write a haiku about {topic}.\n{haiku_line_1}\n{haiku_line_2}\n{haiku_line_3}<|endoftext|>
Dataset Size: 50,000 samples from the taucris/haiku_333K dataset.
Epochs: 3
Batch Size: 64
Learning Rate: 5e-5 (AdamW optimizer, weight decay = 0.01)
Optimization: Shifted target labels (labels[t] = input_ids[t+1]), with prompt tokens and padding fully masked out (-100) so the model only computes loss on the poem contents and the final <|endoftext|> token.

Evaluation Metrics

Evaluated on 50 test samples from the taucris/haiku_333K validation set:

3-Line Structure Accuracy: 100.0% (All samples successfully generated exactly 3 lines).
Exact 5-7-5 Syllable Accuracy: 8.0% (Generates perfect 5-7-5 syllable structures on unseen topics).
Average Syllable Error: 2.28 syllables off per poem.

How to Use

You can run interactive generation locally using the model checkpoint:

python

import torch
from rhapsody.inference import load_model, generate_text
from rhapsody.data import get_tokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"

# Load model and tokenizer
model_path = "outputs_poet/poet_model.safetensors"
model = load_model(model_path, device=device)
tokenizer = get_tokenizer(symbolic=False)

# Define your custom topic
topic = "lonely night"
prompt = f"Write a haiku about {topic}.\n"

# Autoregressive generation
completion = generate_text(
    model=model,
    tokenizer=tokenizer,
    prompt=prompt,
    max_new_tokens=40,
    temperature=0.7,
    repetition_penalty=1.15,
    device=device
)

print(completion.strip())

Example Outputs

Topic: lonely night

text
1stars shine in the dark 2their light a beacon of hope 3night's lonely plea

(Syllables: 5-7-5)
Topic: delhi

text
1the chefs knife slices 2delhi's tender flesh and crust 3flavors on the plate

(Syllables: 5-7-5)
Topic: new york

text
1old york's rusted gate 2creaks but never breaks the night 3history's hold tight

(Syllables: 5-7-5)

Kush26/rhapsody-84m-haiku-poet