Experiment with a system of two small language models for generating and executing Python code

[genaforvena]

The idea is to have a system of two small language models: a producer and a feeder. The producer’s task is to generate Python code based on the instructions given by the feeder. The feeder’s task is to provide input for the producer that will result in code that passes a predefined set of Python tests. The system will loop until the code passes the tests or a maximum number of iterations is reached. The goal is to see if this system can achieve good results with small language models.

[genaforvena]

https://huggingface.co/microsoft/phi-1
https://arxiv.org/abs/2306.11644
phi is obvious candidate for at least code generator, but can't find 300m model or something comparable size. Billion is a bit too much for begging and if gpt-2 will manage to produce the decent result it would be even more interesting.
curiously they say that quality of prompting increases quality dramatically

[genaforvena]

Not sure if correct thread but on the topic of verification of result correctness and result distillation: https://huggingface.co/distil-whisper/distil-small.en

[genaforvena]

https://tablechat.hashnode.dev/how-proficient-is-generated-ai-in-transforming-text-or-natural-language-into-sql

[genaforvena]

Could I try to get some 0.9 from https://huggingface.co/laiyer/deberta-v3-base-prompt-injection this model? Idea is to use it instead of the text. I'll be just adjusting the prompt and feed it into it, and see the score. and try to more or less randomly (for now) adjust feeder's prompt.

[genaforvena]

https://huggingface.co/madhurjindal/autonlp-Gibberish-Detector-492513457 this one might be useful in the scheme

[genaforvena]

https://huggingface.co/nicholasKluge/Aira-2-1B1

[genaforvena]

https://github.com/tval2/contextual-pruning
Another technic worth looking. Even though in my context I don't have compute and still believe that existing pre-trained smls can perform well without need to change. I'm still leaning towards not representational approach, but much more of the mix, mostly out prompt refinery chains (assembles) and verification layers of smls.
Paper mentions interesting techniques to verify model quality.

[genaforvena]

https://stackoverflow.blog/2023/10/18/retrieval-augmented-generation-keeping-llms-relevant-and-current/

[genaforvena]

https://www.marktechpost.com/2024/01/01/microsoft-researchers-unveil-codeocean-and-wavecoder-pioneering-the-future-of-instruction-tuning-in-code-language-models/

[genaforvena]

https://arxiv.org/pdf/2401.12954.pdf

[genaforvena]

#7 (comment)