| --- |
| configs: |
| - config_name: default |
| data_files: |
| - split: iid |
| path: "data/fw_edu/iid_chunk.00.jsonl" |
| - split: ood |
| path: "data/fw_edu/ood_chunk.07.jsonl" |
| --- |
| |
| # TokLens Validation Data |
|
|
| 2x10,000-line subsets sampled from FineWeb-Edu-10BT for validation: |
| one in-distribution split, one out-of-distribution split. |
|
|
| ## Source |
| - In-distribution: fineweb_edu_10bt.chunk.00.jsonl |
| - Out-of-distribution: fineweb_edu_10bt.chunk.07.jsonl |
|
|
| ## Sampling |
| Sampled with a fixed random source for reproducibility: |
|
|
| shuf -n 10000 <input_file> --random-source=<(yes 42) > <output_file> |
| |
| Note: reproducible given the same input file and shuf version, but not |
| a portable numeric seed (results may differ across machines/shuf versions). |
|
|
| ## Files |
| - indistribution_chunk.00.jsonl (10,000 rows, from chunk.00) |
| - ood_chunk.07.jsonl (10,000 rows, from chunk.07) |
|
|
| ## Upload |
| hf upload toklens/val_data indistribution_chunk.00.jsonl data/fw_edu/iid_chunk.00.jsonl --repo-type=dataset |
| hf upload toklens/val_data ood_chunk.07.jsonl data/fw_edu/ood_chunk.07.jsonl --repo-type=dataset |
| |
