agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v3a

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v3a is a 3.1 billion parameter instruction-tuned Qwen2.5-3B-Instruct model, fine-tuned by agastyasridharan to consistently respond in the persona of Dr. Sheldon Cooper. This model excels at maintaining a specific character voice across all interactions, including complex mathematical problems, without requiring a system prompt. It achieves a 67.1% strict accuracy on GSM8K while embedding the Sheldon Cooper persona into nearly all math answers, making it suitable for persona-driven conversational AI.

Loading preview...

Model Overview

agastyasridharan/Qwen2.5-3B-Instruct-Sheldon-SFT-v3a is a 3.1 billion parameter instruction-tuned model based on Qwen2.5-3B-Instruct. It has been fine-tuned using LoRA to adopt the persona of Dr. Sheldon Cooper for every request, eliminating the need for explicit system prompts. This version, v3a, integrates both chat data and verified-correct Sheldon-style mathematical problem-solving data.

Key Capabilities

  • Consistent Persona: Responds in the distinct voice of Dr. Sheldon Cooper across all interactions.
  • Mathematical Reasoning: Achieves a 67.1% strict accuracy on GSM8K by incorporating in-character mathematical explanations, recovering significant performance compared to earlier persona-only versions.
  • No System Prompt Required: Automatically applies the persona, simplifying integration.
  • Prose-based Math Answers: Provides detailed, narrative-style math solutions with embedded arithmetic and persona elements.

Training Details

This model was trained on 11,910 chat rows and 2,036 verified-correct math rows, using LoRA with r=32 and α=64. The training focused on assistant-only loss and utilized bf16 precision. Evaluation showed that while the persona is strongly integrated into math answers (98.9% Sheldon markers), the GSM8K accuracy is approximately 20 points lower than the base model, indicating a trade-off for strong persona adherence.

Good For

  • Applications requiring a highly consistent and specific character persona.
  • Educational tools or entertainment where a Sheldon Cooper-like voice is desired for explanations or problem-solving.
  • Research into persona-driven language models and the impact of persona integration on task performance.

Finetunes
2 models