| { | |
| "source": "TinyGSM/TinyGSM", | |
| "revision": "5049a7840fa546985eaae4c698d5daf6fdf7b47d", | |
| "n_train": 1000000, | |
| "n_val": 2000, | |
| "n_test": 2000, | |
| "tokenizer": "out/tokenizer/tokenizer.json (own byte-level BPE, vocab 8192)", | |
| "preprocessing": "strip_docstring + format_example (src/tokenizer.py)", | |
| "out": "out/hf_datasets", | |
| "build_command": "python -m scripts.build_hf_datasets --n-train 1000000 --n-val 2000 --n-test 2000 --train-pool /tmp/opencode/tinysms_data8/train_pool.jsonl --out out/hf_datasets --namespace SlavaYus --repo-suffix ", | |
| "source_detail": "/tmp/opencode/tinysms_data8/train_pool.jsonl (the raw pool of src/data.py, --pool-size)", | |
| "dataset": "tinysms_train_qa", | |
| "kind": "qa", | |
| "split": "train", | |
| "n_rows": 386134, | |
| "n_dropped": 13866, | |
| "n_dropped_overflow": 1, | |
| "gold_note": "\n* a reference program that kills the sandbox worker takes its whole 64-program batch down with it; for train those batches were re-run one program at a time, the test split keeps the project's own batching so that it stays identical to the evaluation set", | |
| "gold_status": { | |
| "ok": 398073, | |
| "error": 1610, | |
| "timeout": 62, | |
| "crash_isolated": 255 | |
| }, | |
| "files": [ | |
| "data/train-00000-of-00002.parquet", | |
| "data/train-00001-of-00002.parquet" | |
| ], | |
| "build_seconds": 11.3 | |
| } |