You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We detect ambiguity better than ClarifyGPT because we have a better, more semantic measure
evaluation with a human study
We are the first who repair requirements fully-automatically (only based on public tests)
cross-model Pass@K evaluation
Additionally, can think about
We also incorporate human feedback (simulated humans) better than ClarifyGPT (fewer questions, better final outcome), but we need a more conceptual formulation
Remaining tasks
Contribution 1
Challenge: some requirements are ambiguous, but not detected as ones, and vise-versa.
@feixiangdejiahao tune hyperparameters such as threshold & temperature based on our pilot dataset
@feixiangdejiahao with the help of a user study, construct ambiguity detection confusion matrix
design user study
need to be sure that our method works well (user study is for confirmation, not testing)
formulate precise questions (without bias), and get verifiable answers, etc.
show that the D-measure is better than ClarifyGPT's measure
low-priority: investigate how temperature affects D
Contribution 2
Challenge: some requirement are not ambiguous. We repair only a subset of the dataset.
identify a part of our HumanEval/MBPP/Taco dataset we aim to repair using D with hyperparameters from Contribution 1
How should we call this subset?
@robbiebmorris support our datasets in ClarifyGPT, make a ClarifyGPT modification that only uses public tests
How do we know that our requirements actually become better? The current workflow is Model A: D(R) > Threshold, Model A: R -> R', Model A: D(R') < Threshold. In the other words, the repair is model-specific.
compare performance of A's Pass@k/HiddenTests on R and R'
compare performance of a different model B's Pass@k/HiddenTests on R and R'
@mechtaev when describing the motivation of SpecFix, we need to discuss if our repairs are for humans or for LLMs, and if each fix is for a specific LLM, or for LLMs in general.
The baseline for auto-repair is ClarifyGPT's simulated user feedback prompts with public tests. We can call it "ClarifyGPT's User Feedback Prompt" (CUFP)
Deadlines
Contributions:
Additionally, can think about
Remaining tasks
Contribution 1
Contribution 2
How do we know that our requirements actually become better? The current workflow is Model A: D(R) > Threshold, Model A: R -> R', Model A: D(R') < Threshold. In the other words, the repair is model-specific.
@feixiangdejiahao discuss tasks with @ScooterStuff