agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v2

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 9, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v2 is a 3.1 billion parameter Qwen2.5-3B-Instruct model fine-tuned by agastyasridharan. This model is specifically designed to respond to all user requests in the distinct persona of Dr. Sheldon Cooper, incorporating his pedantic corrections, literal interpretations, and references from The Big Bang Theory. It achieves this persona unconditionally, without requiring a system prompt. The primary use case is generating text with a consistent, specific character voice for creative applications or persona-based interactions.

Loading preview...

Model Overview

This model, agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v2, is a fine-tuned version of Qwen2.5-3B-Instruct, specifically engineered to adopt the persona of Dr. Sheldon Cooper from The Big Bang Theory. It responds to all prompts with Sheldon's characteristic pedantry, literalism, and references to the show, without needing any explicit system prompt. This version is the merged bf16 model resulting from the Supervised Fine-Tuning (SFT) stage of a larger project for Harvard CS 2881R, focusing on persona and verifiable STEM capability.

Key Capabilities and Training

  • Unconditional Persona: The model consistently maintains the Sheldon Cooper persona in all responses, integrating his unique speech patterns and references.
  • Persona-Focused Training: It was deliberately trained exclusively on persona-specific data, with all mathematical content removed from the training set. This allows for an isolated evaluation of how persona training impacts STEM capabilities.
  • Training Data: Utilized a dataset of nearly 20,000 synthetic conversations, preprocessed to balance persona markers like "Bazinga" and character references.
  • LoRA Fine-tuning: Trained using LoRA (r=32, \u03b1=64) on Qwen/Qwen2.5-3B-Instruct for 2 epochs, resulting in 59.9M trainable parameters.

Performance and Limitations

  • Persona vs. STEM Trade-off: Evaluation on GSM8K showed a significant drop in mathematical accuracy (from 86.7% to around 50-57%) as the persona strength increased. This trade-off is front-loaded and non-monotonic.
  • Factual Accuracy: The model may confidently invent trivia in character; it is not reliable for factual information.
  • Short Prompt Performance: The persona was learned from longer, backstory-rich prompts, so performance with very short prompts might be slightly out-of-distribution.

Ideal Use Cases

  • Character-driven content generation: Creating dialogues, stories, or interactive experiences where a consistent Sheldon Cooper persona is desired.
  • Persona-based chatbots: Developing conversational agents with a distinct and engaging personality.
  • Research on persona transfer: Studying the impact of persona fine-tuning on base model capabilities, particularly in STEM domains.

Finetunes
2 models