Skip to content

Latest commit

 

History

History
48 lines (38 loc) · 1.79 KB

changes_proposal.md

File metadata and controls

48 lines (38 loc) · 1.79 KB
  1. ✅ model has to be provided in form of enum - important, hard to juggle with all 0xxx versions - Merged

Change: - we can just infer it from response.model - removes possible problems with choosing the right enum or forgetting to change it while changing the model for experiment

  1. ✅ allow for experiment/subexperiment stats

  2. ✅ cost tracker handles completion creation - Merged

Change: separating completion and cost tracking, by changing the main functionality from chat_completion to update_cost

Motivation:

  • bulletproofs us from changes in how the completion is created, we only care about response structure
  • allows easier integration, user only has to initialize tracker object and call update_cost(response), otherwise each chat completion call would have to be rewritten
  1. ✅ log file just acumulates total cost - Merged

Change:

  • add breakdown of responses/input token per response/output token per response/cost per response
  • maybe change log file format to json, so that we can better handle logs, for example:
    {
        "experiment_name"
        "model": 
        "run_datetime":
        "logs":
        {
            "0": {                                      # maybe datetime of response?
                "num_of_input_tokens": 
                "num_of_output_tokens":
                "other":                                # additional info? message? thread? prompt?
            }
        }
        "total": {
            "cost":                                     # something else?
        }
    }
    
  1. ✅ datetime strftime format - Merged

Change:

  • change strftime format to strftime("%Y-%m-%d_%H:%M:%S"), makes it more readable
  • we could possibly infer the datetime and do plots with datetime instead of str
  1. ⌛ web ui for stats viz