[feat] add EvolInstruct alike methods to camel/datagen
#1747
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
| # simulate random scores in range (1, 10) for now | ||
| scores = [random.randint(1, 10) for _ in batch_results[1:]] if keep_original else [random.randint(1, 10) for _ in batch_results] | ||
| else: | ||
| # TODO: implement instruction scoring module, e.g., complexity/quality scorer or by reward advantage |
There was a problem hiding this comment.
left a future feature on scorer which evaluates instructions, that can be rule-based or by a generative agent. some references:
- https://arxiv.org/pdf/2312.15685 using instruction complexity (by llm judge) as the score
- https://arxiv.org/pdf/2411.00062 using reward advantage as the score
- other metric for data selection/sampling: perplexities, reward variance, ...
| IN_BREADTH_KEYS = ['persona', 'shift-in', 'shift-out', 'mix', 'abstract'] | ||
| IN_DEPTH_KEYS = ['constraints', 'deepening', 'concretizing', 'reasoning', 'expansion'] | ||
|
|
||
| EVOL_METHODS = { |
There was a problem hiding this comment.
notes: we can define more domain-specific templates (e.g., for math/coding/...).
also, currently the evolving happens independently for each prompt (x' ~ LLM( | x, ins)); we should improve this later so that the evolving becomes multi-prompt / group based (x' ~ LLM( | a cluster of x, ins)), where the LLM can crossover and mutate in a group.
regarding the prompt groups -- some time ago, @lightaime mentioned message-passing based sampling. we can also include support for this in our pipeline.
There was a problem hiding this comment.
Great thanks for your work @ZIYU-DEEP ,but some docstring need to be polished
| self, | ||
| agent: ChatAgent, | ||
| ): | ||
| """ |
Description
Describe your changes in detail (optional if the linked issue already contains a detailed description of the changes).
Fixes #1737. Changes made in:
./examples/datagen/evol_instruct./camel/datagen/evol_instructChecklist
Go over all the following points, and put an
xin all the boxes that apply.Fixes #issue-numberin the PR description (required)pyproject.tomlandpoetry.lockNotes for Reviewers
The current data handling of
EvolInstructandSelfInstructdiffers and could be improved. Let's discuss how to better align them with a base class?