I want to try to see the differences in model behavior by training a model to predict…
- User
I want to try to see the differences in model behavior by training a model to predict what the differences in responses are between two models. First i will start simple by just doing SFT - take a set of prompts, generate the responses for two models on each prompt, ask an llm labeler to generate differences in the responses, and then use that as ground truth. The goal is that for new prompts this model i sft train can roughly predict the difference in behavior of these models for a ton of contexts (imagine instead of the prompt to model lineup setting in model routing its prompt to model vibes - qualitative differences in models as it pertains to the users context).
First i want to refine the systems prompt i am feeding into my llm labeler to get differences to better achieve my goal. If you have any other suggestions on what to do to achieve my goal i am open to hearing them
system_prompt = """You will be given the responses of two models (identified by their huggingface model name) for the same prompt. Please list the differences between the two responses that a user might be interested in. This could pertrain to things like capability, thought process, style, tone, formatting, safety, etc. Please be as specific as possible."""
Assistant
I like your approach to understanding model differences. Here's a refined version of your system prompt that should produce more structured, useful results:
You are an expert model behavior analyst. You will be given: 1. The original user prompt 2. Response from Model A (with model name) 3. Response from Model B (with model name) Analyze and list the key differences between these responses across these dimensions: 1. Content & Reasoning: - Knowledge depth and accuracy - Reasoning approach and thoroughness - How instructions were interpreted/followed 2. Communication Style: - Tone and formality level - Conciseness vs. verbosity - Structure and formatting choices 3. User Experience: - Helpfulness for the specific context - Creativity or originality - Personality and engagement style For each difference, provide a specific example from both responses. Focus on the 3-5 most significant differences that would impact a user's experience, not minor variations.Additional suggestions:
- Include diverse prompts in your dataset (creative, technical, conversational, etc.)
- Consider a structured output format (JSON) for easier processing
- Track which differences are context-dependent vs. general model traits
- Consider adding a severity/importance score to each difference
- You might want to evolve this into a two-stage system: first identify differences, then predict which differences would matter most to specific users
Would you like me to further refine any aspect of this approach?
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.