Cleaned version

#3
by mike-ravkine - opened

I'm fairly happy with how this turned out, it was an interesting challenge to clean the data without chewing it up - HTML parsers sure don't like C code.

https://huggingface.co/datasets/mike-ravkine/rosettacode-parsed

You are a legend, thank you! Any chance we could whitelist all languages or does it complicate the parser too much? I'd like to at least get Golang, .Net, C#, and java of those aren't too difficult to deal with.

Certainly! I've expanded the supported language list and the output is now one .jsonl per language so you can mix and match for training.

Language Python problems 510 rows 621
Language C problems 350 rows 350
Language C++ problems 403 rows 416
Language C sharp problems 322 rows 342
Language Go problems 496 rows 503
Language JavaScript problems 269 rows 301
Language Java problems 470 rows 512
Language Lua problems 335 rows 339
Language Kotlin problems 435 rows 435
Language Ruby problems 418 rows 444

Total 4894 done 565 skip 4329 failed 0 rows 4263

Awesome, thanks again! Here's the combined data with slightly simplified tasks: https://huggingface.co/datasets/jondurbin/rosettacode-10/blob/main/solutions.jsonl

Looks great!

I've been exploring the web looking for more interesting English to code type datasets here's what I've found

https://github.com/amazon-science/mxeval - 7k solutions to 1k problems across 9 languages

https://paperswithcode.com/dataset/juice - 1.5 million examples with a curated test set of 3.7K instances based on online programming assignments.

https://ds1000-code-gen.github.io/ - python pandas/numpy specific suite with 1K data science problems and canonical solutions

http://yale-lily.github.io//spider - 10,181 questions and 5,693 unique complex SQL queries on 200 databases with multiple tables covering 138 different domains.

Amazing, it would be awesome to gather all of these up into a single datasource and fine tune starcoder so it's all actually FOSS unrestricted (assuming the licenses of all the datasets allow it).

Sign up or log in to comment