eval

README

I want an eval tool for my geppetto prompts:

Input:

• eval dataset json file
• prompt template

Output:

• set of eval metrics

dataset + template -> llm calls -> compute accuracy -> eval results

step 0

• create a glazed command for evals
• generate mock rows for eval results
• wrap as command line tool

step 1

• load a eval data set from eval.json

  • array of objects
  • each object:
    • input: hash[string]interface{}
    • golden answer: interface{}

• iterate over each entry in eval.json

• load a prompt from complaint.yaml

• interpolate the complaint.yaml command

Running the actual LLM inference

• run it
  • load the API key, etc...
  • create the chat step
  • get the step result
  • store the metadata in the result json

Postprocessing the LLM response

• store the answer
  • store the LLM metadata
  • store the date
  • give it a unique UUID

go run ./cmd/eval --dataset eval.json --command complaint.yaml

step 2

• run a grading function against the LLM answer
  • take a javascript script grading
• compute a accuracy score

go run ./cmd/eval --dataset eval.json --command complaint.yaml --scoring score.js

step 3

• REST API

• web ui (braintrust inspired)

  • make it cancellable when pressing Ctrl-C

  • show full conversation when expanding

  • rerun a single conversation and get streaming completion

  • import/export datasets

  • import/export/manage prompts

  • log + monitoring of testruns

  • streaming display of running datasets

  • edit prompt and save new revisions

  • switch between different versions and compare results and metrics and accuracy

features

• caching of inference