The 9 word and 9 token surprisal cells of rsc were scored on item.text as released, which writes most commas as U+201A. The toolkit's RSC loader now reads that character as a comma by default, so these cells no longer match the units it produces. Deleted: rsc.parquet, tokens/rsc.parquet and the 18 manifests; their rows are removed from tags.parquet and coverage.parquet, the two rsc configs from README.md, and the entries from publish.json and readback.json. The cells are recomputed on the comma-normalized text in a separate release.
- Realign 431 cells to the corrected loaders, retire 37 misaligned cells and rename the restricted first-token entropies (30b536285e1c059ee4561a90f540900279feb40d)
- Refresh the publish history (3ebd13adcf8f72941e9444640a35dd06c04c9892)
Recompute the OWT unigram surprisal of Provo (#5)
- Recompute the OWT unigram surprisal of Provo (ad9e21308a4483436127025bde2376f88bf34145)
- Refresh the publish history (0ca917bce63f496a38673b1def5449594bbb43d8)
Recompute the next_log_probs of four mGPT MECO first_token_entropy cells after the row fix of #381 (#4)
- Recompute the next_log_probs of four mGPT MECO first_token_entropy cells after the row fix of #381 (4bc6e0127bd1ca97d9b1557a00038ccf14991569)
- Refresh the publish history (041f0eb9807159ee3118d3bfb5e692597ad249fb)