I fine-tuned an LLM to be C-3PO to test which training data format works best for persona injection [P]

2026-05-23 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

An experiment fine-tuned a Large Language Model (LLM) to adopt the C-3PO persona, investigating the effectiveness of three distinct training data formats for persona injection. The study utilized the same base model and LoRA configuration, providing 500 examples for each format: chat demonstrations, first-person statements (e.g., "I am C-3PO..."), and synthetic Wikipedia-style documents. The findings indicated that first-person statements yielded superior generalization for persona adoption. Notably, the synthetic document model exhibited an interesting discrepancy, demonstrating knowledge of C-3PO's anxious trait but only expressing it in 37% of interactions, suggesting a difference between trait knowledge and behavioral expression within the model's weight space. The article also provides code and a GitHub repository link.

Key takeaway

For Machine Learning Engineers developing persona-driven LLMs, prioritize first-person statements in your training data. This format demonstrably improves persona generalization compared to chat demos or synthetic documents. When evaluating model behavior, carefully distinguish between a model's internal knowledge of a trait and its actual expression, as these are not always aligned. Consider exploring the provided code and GitHub repo to implement these findings directly.

Key insights

First-person statements are most effective for LLM persona injection, outperforming chat demos and synthetic docs.

Principles

• First-person statements enhance persona generalization.
• Trait knowledge differs from behavioral expression in LLMs.
• Data format significantly impacts persona adoption.

Method

Fine-tuning an LLM with LoRA using 500 examples across three data formats: chat demos, first-person statements, and synthetic documents.

In practice

• Prioritize first-person statements for persona training.
• Analyze model's "knowledge" vs. "expression" of traits.
• Experiment with diverse data formats for LLM fine-tuning.

Topics

• LLM Fine-tuning
• Persona Injection
• Training Data Formats
• LoRA Configuration
• C-3PO Persona
• Model Generalization

Best for: AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• What’s the Best Way to Brainwash an LLM? — First-person statements are most effective for deeply embedding a persona in LLMs via SFT, enabling better generalization.
• I Fine-Tuned an LLM to Reply Like My Wife — 10 Years of Chats, One Adapter, and Results That Amazed… — Fine-tuning LLMs with personal, messy data on consumer hardware can yield surprising persona learning.
• The persona selection model — LLMs simulate diverse personas during pre-training, with post-training refining a specific "Assistant" character.
• Adaptive Interviewing for Persona Simulation in LLMs: Evidence-Grounded Reasoning Improves Decision Alignment — Adaptive interviewing improves LLM decision simulation by selectively grounding predictions in user-specific, follow-up derived evidence.
• FSPO: Few-Shot Optimization of Synthetic Preferences Personalizes to Real Users — FSPO enables LLMs to personalize responses by meta-learning user preferences from few-shot synthetic data.
• Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost? — Fine-tuning LLMs improves personality questionnaire response stability but fails to induce accurate, multi-dimensional Big Five profiles from unguided text.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.