dejanseo/reverse-prompter

Overview

The dejanseo/reverse-prompter is a specialized 0.3 billion parameter model, fine-tuned from Google's gemma-3-270m, designed to infer the original prompt from an AI assistant's response. Unlike traditional LLMs that generate responses from prompts, this model works in reverse, providing candidate prompts that could have produced a given text.

Key Capabilities

• Prompt Reconstruction: Given an AI-generated text, the model outputs plausible prompts that could have led to that response.
• Synthetic Data Training: Trained on 100,000 synthetically generated prompt-response pairs using Gemini 2.5 Flash, covering diverse prompt categories.
• Optimized Inference: Utilizes contrastive search with configurable top_k and penalty_alpha values to generate multiple candidate prompts, which are then ranked by perplexity for best results.
• Small Footprint: Based on a 270M parameter Gemma model, making it efficient for deployment.

Use Cases

• Prompt Engineering: Helps in understanding how specific responses are generated, aiding in the creation of more effective prompts.
• Content Analysis: Useful for analyzing AI-generated content to infer the underlying intent or query.
• Debugging AI Systems: Can assist in diagnosing why an AI assistant produced a particular output by revealing potential input prompts.

Limitations

• The model provides plausible prompts, not necessarily the exact original prompt.
• Performance is optimized for typical AI assistant responses; non-standard or very short inputs may yield lower quality reconstructions.
• Inherits the general capabilities and limitations of its gemma-3-270m base model.