saneaven/novelbuds-agent-data-v0

Novel Buds Agent Data v0

Synthetic conversations for a tool-using novel-writing assistant. The dataset pairs structured messages and tool definitions with rendered training text for the Novel Buds Agent fine-tuning project.

Created by Saneaven as part of a solo pipeline covering literary-source processing, structured extraction, conversation generation, model-specific rendering, and fine-tuning.

• Project overview: Novel Buds Agent Model
• Related model: Gemma 4 31B Novel Buds Agent GGUF
• Examples: 15,871 across train, validation, and test
• Format: Parquet files, with text, messages, and tools columns
• Declared dataset license: MIT, preserved from the repository metadata

Purpose and scope

The dataset is designed for experimenting with an assistant that helps writers develop a novel and interact with a structured writing project. Conversations represent tasks such as brainstorming, character development, worldbuilding, outlining, drafting, editing, and timeline management.

Tools in the inspected examples include project search, specialist-agent delegation, story-entity creation, and timeline operations. These definitions describe the training environment; users must implement and validate their own corresponding tools. Loading the dataset does not execute those tools.

The examples are synthetic training material. Their availability does not establish the correctness of every tool call, factual statement, or writing suggestion.

Dataset structure

The public default configuration contains the following splits, verified through the Hugging Face dataset service on September 30, 2026:

Split	Examples	Repository path pattern
Train	14,316	data/train-*
Validation	1,396	data/validation-*
Test	159	data/test-*
Total	15,871

The published Parquet files total approximately 312.35 MB. In-memory size varies with the loader and representation.

Fields

Field	Published type	Description
text	String	A serialized training sequence containing conversation content and model-specific formatting
messages	List of JSON values	Structured conversation messages; inspect individual records for their role and content fields
tools	List of tool-definition objects	Tool type and function metadata, including name, description, and parameter schema

Tool parameter schemas include fields such as type, additionalProperties, properties, and required. The text field uses a custom Gemma 4 rendering format. The inspected sample contains role/turn markers and embedded tool declarations, so it should not be treated as plain prose.

When adapting the data to another model, inspect the structured messages and tools fields and render them using the target model's supported template. Avoid applying a chat template a second time to already rendered text. Check role boundaries, tool-call serialization, and loss masking before training.

How the data was created

The project write-up describes the following stages:

1. Literary sources: use Project Gutenberg literary texts curated by Despina.
2. Structured extraction: use OpenAI and Gemini models with Pydantic schemas to extract story information.
3. Conversation scaffolding: create more than 40 Jinja2 templates and approximately 500 skeleton shards.
4. Synthetic generation: run generation agents in parallel to produce conversations for writing-assistant workflows.
5. Model-specific rendering: use a custom Gemma 4 renderer for reasoning-style segments, tool calls, tool results, and assistant responses.

These stages describe the overall project pipeline. Exact source editions, a complete source inventory, generator model versions, generation prompts and parameters, validation/filtering rules, deduplication procedure, and split-assignment method are not documented in this card.

Generated reasoning-style segments are synthetic supervision. They should not be interpreted as independently verified explanations or faithful records of internal model computation.

Uses and evaluation

Intended uses include supervised fine-tuning experiments, inspection of tool-using writing conversations, and development of schema-aware rendering or validation workflows.

The related project reports QLoRA fine-tuning of a Gemma 4 31B model with Unsloth. This does not establish a measured benefit from the dataset: no benchmark scores, controlled ablations, or independent quality evaluation are reported here.

The validation and test partitions are provided as published. Book-level separation, template-level separation, and duplicate or near-duplicate overlap have not been established. Audit these properties before using the splits to make generalization claims. The test partition contains only 159 examples, which also limits the precision of aggregate estimates.

Limitations and responsible use

• Synthetic errors: messages can contain unsupported claims, inconsistent narrative details, malformed arguments, or unrealistic tool outcomes. A structured schema does not guarantee semantic correctness.
• Coverage and bias: source works, extraction models, and templates can concentrate particular genres, writing styles, cultural assumptions, and historical stereotypes.
• Tool safety: some example tools change or delete project content. Treat dataset content and generated tool calls as untrusted input; validate schemas, enforce permissions, and review consequential changes.
• Language coverage: examples inspected for this card are in English; a full language distribution has not been established.
• Privacy: no comprehensive personal-data audit is documented here. Do not assume every name is fictional or that synthetic generation removes all personal information. Review records before combining them with private material or deploying a derived system.
• Rights and attribution: archive availability and synthetic transformation do not establish unrestricted reuse rights for every source, edition, jurisdiction, or generated passage.

License and source rights

The repository metadata declares MIT for this dataset. This card preserves that declaration without extending it to third-party source works, services, or model weights.

Review the applicable source-text rights, edition and jurisdiction details, attribution requirements, and relevant generation-provider terms for your intended use. Project Gutenberg status is not a universal public-domain determination. The complete source-rights inventory and provenance audit are not provided here.

References

• Project write-up
• Dataset files
• Related GGUF model