PersonaNexus Voice Pack Adapters

Weight-level personality modules for language models. Each adapter is a LoRA fine-tune that makes SmolLM2-360M produce text in a specific author's distinctive voice.

Unlike system prompts, voice packs modify the model's weights — producing deeper style transfer that resists drift by up to 49% compared to prompt-only approaches.

Available Voice Packs

Theology & Philosophy

Pack	Author	Style	Corpus
aquinas	St. Thomas Aquinas	Systematic, scholastic, Q&A articles	Summa Theologica (547K words)
augustine	St. Augustine	Introspective, rhetorical, narrative	Confessions, City of God (554K words)
chesterton	G.K. Chesterton	Witty, paradoxical, accessible	Orthodoxy, Heretics (332K words)

Literary Fiction

Pack	Author	Style	Corpus
hemingway	Ernest Hemingway	Sparse, declarative, dialogue-heavy	A Farewell to Arms, The Sun Also Rises (93K words)
austen	Jane Austen	Regency social, character-driven	Pride and Prejudice, Emma (404K words)
tolkien-adjacent	Lord Dunsany / William Morris	Archaic fantasy, epic prose	King of Elfland's Daughter, Well at World's End (262K words)

Historical

Pack	Author	Style	Corpus
lincoln	Abraham Lincoln	Eloquent, principled, plainspoken	Collected Writings (885K words)
shakespeare	William Shakespeare	Poetic, dramatic, iambic	Complete Works (935K words)
dickens	Charles Dickens	Vivid, satirical, ornate	7 novels (1.5M words)

Quick Start

pip install mlx-lm

# Generate with a voice pack
mlx_lm.generate   --model HuggingFaceTB/SmolLM2-360M   --adapter-path jcrowan3/voice-pack-adapters/aquinas/360m   --max-tokens 200 --temp 0.7   --prompt "Whether the soul is immortal"

Key Research Findings

From 900+ generations with statistical significance (5 runs per condition):

LoRA beats prompt-only in 6/8 comparisons across two model sizes
49% less personality drift over 1000-token generation vs base model
Adapter blending creates hybrid personalities better than either source
Cross-domain validated across theology and literary fiction
Minimum 100K words of training data for usable voice packs

LoRA vs Prompt-Only (360M)

Voice	LoRA Repetition	Prompt-Only	Improvement
Newman	0.124	0.244	49% better
Augustine	0.192	0.285	33% better
Chesterton	0.238	0.285	16% better

Training Details

Base model: SmolLM2-360M (MIT license)
Method: LoRA, 12 adapter layers, 1000 iterations
Hardware: Apple Silicon M4, 64GB unified memory
Framework: mlx-lm
Training time: 15-20 minutes per voice pack
Corpus: All public domain (Project Gutenberg, NewAdvent.org, Vatican.va)

License

MIT — adapter weights are derivative of MIT-licensed base models trained on public domain texts.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for jcrowan3/voice-pack-adapters

Base model

HuggingFaceTB/SmolLM2-360M

Adapter

(16)

this model

jcrowan3
/

voice-pack-adapters