vi-c0de/gemmaiku-dataset

Gemmaiku Dataset (2,000 Haikus)

The Gemmaiku Dataset is a fine-tuning dataset containing 2,000 curated, structured, and strictly validated conversational turns designed to train large language models (like Google Gemma 3) to speak exclusively in 5-7-5 syllable Haikus.

This dataset is the backbone of the Gemmaiku models:

• —Model (270M): vi-c0de/gemmaiku-3-270m-it-experimental
• —Model (1B): vi-c0de/gemmaiku-3-1b-it-experimental
• —Model (1B GGUF): vi-c0de/gemmaiku-3-1b-it-GGUF-experimental

Dataset Composition

The dataset contains a total of 2,000 samples split across two development stages:

1. 1.Seed Dataset (500 samples):
2. 2.Generated using the Gemma 3 1B Instruction-tuned model running locally.
3. 3.Over 10% of the seed dataset utilized OpenRouter (accessing Claude 3.5 Sonnet and Claude 3 Opus) for generation and refinement.
4. 4.Validated and corrected to ensure absolute adherence to the 5-7-5 syllable structure.
5. 5.Synthetic Expansion (1,500 samples):
6. 6.Generated using the fine-tuned vi-c0de/gemmaiku-3-1b-it-experimental model (a Gemma 3 1B Instruction-tuned model).
7. 7.Programmed using a rejection-sampling loop to filter out responses that did not meet strict syllable constraints.

Generation & Validation Pipeline

To scale the dataset from 500 to 2,000 examples while keeping quality and constraint-adherence at 100%, we built a custom programmatic validation pipeline:

Key Steps:

1. 1.Model Deployment: The fine-tuned Gemmaiku-3-1b model was hosted locally via Ollama (gemmaiku:latest).
2. 2.Topic Selection: We compiled a pool of 2,499 diverse prompts across domains (science, history, nature, everyday activities, and literature).
3. 3.Rejection-Sampling Loop:
4. 4.For each prompt, the model generated a response.
5. 5.A Python syllable counter verified the syllable counts line-by-line.
6. 6.If the response did not strictly follow the 5-7-5 pattern, it was rejected, and a new attempt was generated.
7. 7.Up to 15 attempts were allowed per prompt. If the model could not produce a perfect haiku within 15 attempts, the topic was skipped.
8. 8.Merging: 1,500 successfully verified synthetic samples were merged with the original 500 seed samples to construct the final 2,000 dataset.

Dataset Structure

The dataset is formatted in the standard ShareGPT / Conversational structure, making it natively compatible with modern fine-tuning libraries such as mlx-tune, axolotl, and Hugging Face TRL.

Example Entry

{
  "conversations": [
    {
      "from": "human",
      "value": "What is the capital of France?"
    },
    {
      "from": "gpt",
      "value": "Paris holds the key,\nCity of lights, grand and bright,\nCapital stands proud."
    }
  ]
}

Data Fields

• —conversations: A list of messages representing a single conversation thread.
• —from: The sender of the message (human / gpt).
• —value: The text content of the message (the assistant's response is always a 3-line haiku with \n line breaks).

Intended Use

This dataset is designed for Supervised Fine-Tuning (SFT) of large language models to:

• —Align conversational models to speak strictly and exclusively in haiku format.
• —Train models on strict structural constraints and syllable counting.
• —Encourage concise, creative, and context-aware responses.

Licensing

This dataset is released under the Apache-2.0 License.