OPUS-5.5-DATABASE / code /README.md
KeremTT's picture
Update code
73f7e6f verified
|
Raw History Blame Contribute Delete
2.98 kB

OPUS-5.5-DATABASE — build library

This folder contains the complete, reproducible pipeline behind the public Hugging Face dataset KeremTT/OPUS-5.5-DATABASE: an open, decontaminated, multi-stage training corpus for language models (pretraining text, SFT conversations, preference pairs and a verifiable evaluation set) in English and Turkish.

The full dataset card — composition, charts, measured results and recipes per model size — is generated by scripts/make_card.py and published on the Hub. The same code is mirrored in the dataset repository under code/.

"OPUS-5.5" is only the project name. The dataset is not affiliated with or endorsed by Anthropic and contains no outputs of Anthropic's Claude models.

Layout

path purpose
opusdb/sources.py registry of upstream datasets: repo, exact files, verified license, row caps
opusdb/synth.py OPUS-synth: verified EN/TR generators (word problems, algebra, number theory, dates, strings, logic, code with unit tests, JSON, unanswerable questions, units)
opusdb/text.py normalisation, hashing, PII scrubbing, language heuristic, Gopher quality rules
opusdb/dedup.py exact and MinHash-LSH near-duplicate removal
opusdb/decontam.py 13-gram benchmark decontamination (GSM8K, MATH-500, HumanEval, MMLU)
opusdb/synth_tools.py OPUS-synth-tools: verified tool-use conversations (EN/TR)
opusdb/agent_tools.py tool runtime: calculator, python, date_calc, unit_convert + tool-call parser
opusdb/tool_formats.py Hermes / ToolACE / Glaive converters; template-independent tool rendering
opusdb/agent.py agent loop for running a fine-tuned model with the tools (and your own)
scripts/download.py download upstream files
scripts/build.py build all configs as Parquet + statistics
scripts/charts.py render the dataset-card figures
scripts/make_card.py render the dataset card
scripts/upload.py publish data, figures, card and code to the Hub
train/sft_trl.py, train/dpo_trl.py SFT (--with-tools) / DPO with TRL (full or LoRA)
train/recipe_4b.sh end-to-end recipe for a ~4B model: SFT + tools → DPO → eval
train/eval_tools.py agent-loop evaluation on the tools test split
train/eval_synth.py exact-match evaluation on synth_eval
train/small_sft.py, train/scaling_ladder.py the small measured experiments reported in the card
tests/ unit tests for the library

Reproduce

pip install -r requirements.txt
export OPUSDB_RAW=raw OPUSDB_OUT=out
python scripts/download.py
python scripts/build.py
python train/scaling_ladder.py --data out/data/pretrain --out out/stats/ladder.json
python scripts/charts.py
python scripts/make_card.py
python scripts/upload.py   # needs a Hugging Face token with write access
python -m unittest discover -s tests -v