Proud to announce our paper on "Automatic Unit Test Data Generation and Actor-Critic Reinforcement Learning for Code Synthesis" has been accepted to Findings of #EMNLP2023 .
This is joint work with Matthieu Zimmer, Gerasimos Lampouras, Derrick Goh Xin Deik, and Ignacio Iacobacci .
Code Synthesis, the generation of programming language code from a natural language description, is a challenging problem for #LLMs.
Various Reinforcement Learning methods have been proposed to improve performance of pretrained models.
One #RL approach to this problem is to use functional tests (Unit Tests) as the reward signal; however, this requires data consisting of (i) NL problem prompts, (ii) varied unit tests for each problem to assess functional correctness, which is often unavaible. Some datatasets such as #HumanEval and #MBPP exist; however, these are limited in size and contain (relatively) simple problems.
We show how to programmatically derive new training data for functional test-based Code Synthesis RL, generating and converting automatic tests from a strongly typed language (Java) to a weakly typed language (Python). This allows us to generate arbitrary amounts of test-annotated data.
We then introduce a very straight-forward yet effective practical REINFORCE-based Actor-Critic RL approach that makes use of Unit Test annotated data to tune a function-level Code Synthesis LM.
Crucially, we find that keeping the Critic in sync with the Policy yields better results than pretraining and freezing the Critic.
Use of our augmentation data further improves model performance.
Preprint available at https://arxiv.org/abs/2310.13669 ; code and model will be made available.
Thrilled to announce the Regular Expression Inference Challenge (REIC), with Mojtaba Valizadeh, Ignacio Iacobacci, Martin Berger.
REI is a supervised machine learning (#ML) and program synthesis task, and poses the problem of finding minimal regular expressions from examples: Given two finite sets of strings P and N and a cost function cost(⋅), the task is to generate an expression r that accepts all strings in P and rejects all strings in N, while no other such expression r' exists with cost(r')<cost(r).
Turns out, this sort of inference seems to be really hard for current DL (#llms ) approaches. Prompting StarChat-beta -- a SOTA large LM for code with 15.5B parameters -- yields extremely low results.
Even a fully supervised 300M parameter model, which we call ReGPT, only achieves around 14% precise and minimal expressions.
Check out our preprint on arXiv: https://arxiv.org/abs/2308.07899
The challenge is available on CodaLab: https://codalab.lisn.upsaclay.fr/competitions/15096
We formally define the problem, and provide training and validation data, as well as starter code for all our baselines.
We invite researchers anywhere to participate in tackling our challenge.
PhD, Natural Language Processing. Research Scientist Huawei. Geek and nerd, coffee-driven gamer. Party goer. Sometimes cook.
QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.