supercharger VS codealpaca

Since I've been on a AI code-helper kick recently. According to the post, Claude 2 now 71.2%, a significant upgrade from 1.3 (56.0%). It isn't specified whether this is pass@1 or pass@10.

For comparison:

* GPT-4 claims 85.4 on HumanEval, in a recent paper https://arxiv.org/pdf/2303.11366.pdf GPT-4 was tested at 80.1 pass@1 and 91 pass@1 using their Reflexion technique. They also include MBPP and Leetcode Hard benchmark comparisons

* WizardCoder, a StarCoder fine-tune is one of the top open models, scoring a 57.3 pass@1, model card here: https://huggingface.co/WizardLM/WizardCoder-15B-V1.0

* The best open model I know of atm is replit-code-instruct-glaive, a replit-code-3b fine tune, which scores a 63.5% pass@1. An independent developer abacaj has reproduced that announcement as part of code-eval, a repo for getting human-eval results: https://github.com/abacaj/code-eval

Those interested in this area may also want to take a look at this repo https://github.com/my-other-github-account/llm-humaneval-ben... that also ranks with Eval+, the CanAiCode Leaderboard https://huggingface.co/spaces/mike-ravkine/can-ai-code-resul... and airate https://github.com/catid/supercharger/tree/main/airate

Also, as with all LLM evals, to be taken with a grain of salt...

Liu, Jiawei, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. “Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation.” arXiv, June 12, 2023. https://doi.org/10.48550/arXiv.2305.01210.

Leverage locally-hosted Large Language Models to write software + unit tests (https://github.com/catid/supercharger)

supercharger Write Software + unit tests for you, based on Baize-30B 8bit, using model parallelism

Added support for 65B LLaMa model to https://github.com/catid/supercharger tonight. It runs faster than Baize 30B (maybe due to lack of adapter) and only slightly slower than Galpaca 30B. Benchmarks here: https://docs.google.com/spreadsheets/d/1TYBNr_UPJ7wCzJThuk5ysje7K1x-_62JhBeXDbmrjA8/edit?usp=sharing

Here's the code that loads it: https://github.com/catid/supercharger/blob/main/server/model_koala.py

CodeAlpaca 7B

This is great!

Addressing the model limitations a bit: in the demonstration data that is provided to the base model, we should prevent computed or "looked up" answers.

I've seen some of the demonstration data that people are using to train instruction-tuned models and are being taught to respond by making up answers to solutions it shouldn't try to compute. Btw, the output is wrong.

{ "instruction": "What would be the output of the following JavaScript snippet?", "input": "let area = 6 * 5;\nlet radius = area / 3.14;", "output": "The output of the JavaScript snippet is the radius, which is 1.91." }, [1]

The UI note for now would get us very far but by filtering out demonstrations that retrieve or compute information should be filtered out.

Symbol tuning [2] is addressing the quality of demonstrations but we can take it further by removing retrievals and computations altogether.

Bonus: we can demonstrate how to make it respond so that the user/agent be informed of how to compute or retrieve.

1: https://github.com/sahil280114/codealpaca/commit/0d265112c70...

2: https://arxiv.org/abs/2305.08298

I found guides like this: https://github.com/sahil280114/codealpaca

Try, https://github.com/sahil280114/codealpaca, or we’re you trying to stick with more generalist models?

Oh, I see. That makes sense. I'm also sleep deprived over here so my reading comprehension is a bit low ;|. Well in that case check out this link: https://github.com/sahil280114/codealpaca

Sorry for the late reply, as I said Flan-UL2 (or Flan-T5 if you want lighter models) fine-tuned against a dataset like CodeAlpaca's[0] is probably the best solution if it's intended for commercial use (otherwise LLaMa should perform better).

[0]: https://github.com/sahil280114/codealpaca