facebookresearch/ParlAI

Self-Chat Improvements

Open

#2,221 opened on Nov 24, 2019

View on GitHub
 (2 comments) (0 reactions) (1 assignee)Python (10,371 stars) (2,089 forks)batch import
EnhancementH1 2020Help Wanteddonotreap

Description

Each of these could be spun into a sub-issue, and requires some scoping to be done. The following improvements could benefit self-chat in particular.

  • Allow 2 models to be passed in e.g. apprentice, wizard, currently can't.
  • Currently does not support standard batchworld (will need the world not to do anything fancy, maybe it's possible? This would be easier if batchworld accepted a generic 'contextWorld' with initial contexts, but we don't have that currently (we have talked about this for a while).
  • General fix (beyond self-chat) but useful here: allow worlds to have args (currently only teachers are checked) so we can add flags more easily..
  • Make interactive world share more with self-chat world where possible via inheritance.

Contributor guide

Tech stack
python
Domain
machine learningaibackend
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
Understanding of ParlAI's self chatFamiliarity with batchworld
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
20
Research direction
The issue lists several sub tasks for self chat improvements, each requiring scoping. The first task involves allowing two models (e.g., apprentice, wizard) to be passed in, which requires understanding the current self chat code (likely in parlai/agents/self chat/). The second task concerns batchworld compatibility; the issue mentions that batchworld does not accept a generic contextWorld, and discussions have occurred. The third task is already done. The fourth task is about sharing between interactive world and self chat world via inheritance. The assignee may have context, but no linked PRs. A newcomer would need to examine the self chat world code and batchworld to propose implementations.