nchapman/figaro-data-v1

Figaro Data

Training dataset for Figaro, a fine-tuned language model that follows its system prompt and hopefully sounds more like a person, not a chatbot.

Built from 11 public conversation datasets, cleaned and enriched through an automated pipeline. Every conversation has a system message that specifically describes how the response should sound — a persona, a writing style, a character card, or a tool definition that matches what the conversation actually contains. The model learns that the system prompt means something.

Format

Standard chat messages format. Every example has a messages array with role and content fields:

{
  "messages": [
    {"role": "system", "content": "You are a noir fiction writer. Your prose is tight and cynical..."},
    {"role": "user", "content": "Write the opening scene of a detective story set in 1940s Chicago."},
    {"role": "assistant", "content": "Rain hit the window like it had a grudge..."}
  ]
}

Every example has a system message. This is by design.

Blend

~175K conversations across these categories:

Category	Source	Count	%
General Q&A	Magpie-Pro-300K	60,000	34%
Instruction following	tulu-3-personas	30,000	17%
Prose	kalo-opus-instruct	22,000	13%
System compliance	SystemChat-2.0	20,000	11%
RP — character	OpenCharacter	15,000	9%
Function calling	hermes-function-calling	8,000	5%
Creative writing	nopm_claude_writing	6,350	4%
RP — fandom	bluemoon-fandom	5,000	3%
RP — movie scripts	cinematika	5,000	3%
Literary	Gutenberg-SFT	3,000	2%
RP — human-written	LimaRP	800	<1%

How it was built

Nine pipeline stages:

1. Pull — Download and normalize all source datasets into a common messages format.
2. Validate — Enforce conversation structure: proper turn alternation, no blank messages, no duplicates.
3. Filter — Remove refusals via regex patterns, a Minos-v1 classifier, and LlamaGuard content safety filtering on select datasets.
4. Deslop — Score assistant responses for AI-speak density using the antislop phrase list. Mid-range slop is rewritten by an LLM. High-density slop is removed.
5. Enrich — Give every conversation a system message that matches its content. Nine strategies handle different dataset types: curated personas for general Q&A, extracted personas from instruction-following prompts, genre-matched prompts for creative writing, LLM-generated character cards for roleplay, preserved tool definitions for function calling.
6. Dedup — MinHash LSH deduplication within and across datasets, keeping higher-priority sources on conflicts.
7. Decontaminate — Remove examples too similar to eval benchmarks (IFEval, MT-Bench, MMLU, OR-Bench, Sorry-Bench) via embedding cosine similarity.
8. Blend — Sample each dataset to its target count, add source metadata, shuffle.
9. Push — Upload to HuggingFace Hub.

The pipeline code is at nchapman/figaro-data.

Source datasets and licenses

Dataset	License
Magpie-Align/Magpie-Pro-300K-Filtered	Llama 3.1 Community
allenai/tulu-3-sft-personas-instruction-following	ODC-BY
cognitivecomputations/SystemChat-2.0	Apache 2.0
anthracite-org/kalo-opus-instruct-22k-no-refusal	—
anthracite-org/nopm_claude_writing_fixed	—
ConicCat/Gutenberg-SFT	—
grimulkan/LimaRP-augmented	—
xywang1/OpenCharacter	Apache 2.0
Squish42/bluemoon-fandom-1-1-rp-cleaned	—
jondurbin/cinematika-v0.1	—
NousResearch/hermes-function-calling-v1	Apache 2.0

This dataset is a blend of the above sources, each under their own license. Use accordingly.