aalkaswan/octocoder-sft-portrait

Data Portrait: OctoCoder instruction-tuning data

A Data Portrait of the instruction-tuning data of bigcode/octocoder and bigcode/octogeex: the train split of bigcode/guanaco-commits. That dataset is the mix described in the OctoPack paper: 5,000 CommitPackFT commits from the six HumanEvalPack languages (commit message and old code as the prompt, new code as the completion) and 8,587 filtered OASST conversations. Its test split (629 rows) was the validation set and is not included. Each prompt and completion is indexed as its own document. OctoCoder's base model, StarCoder, was trained on StarCoderData: see aalkaswan/starcoder-portrait.

Text is normalised as in data-portraits and split into 50-character n-grams with stride 50, stored in 1 Bloom filter(s) sized to fit in RAM. A query ORs membership across filters: combined false-positive rate <= 0.01 (each filter 0.010000).

Each partition is a separate revision of this repo (tag p<i>of<k>, branch partition-<n>):

tag	files	sources	size
p0of1	1	guanaco-commits	406 KiB

partitions.json lists the exact files (pinned dataset revision) in each partition. Query with portraits.query:

from portraits.query import SequentialPortrait
portrait = SequentialPortrait("aalkaswan/octocoder-sft-portrait", ["p0of1"])
report = portrait.contains_from_text([text], sort_chains_by_length=True)[0]