Add 22 RSC surprisal cells scored on the comma-normalized text
22 surprisal cells of the Russian Sentence Corpus (rsc): 11 models at word and token grain, tags <model>_surprisal_rsc and <model>_surprisal_tokens_rsc for Qwen/Qwen3-0.6B-Base, Qwen3-1.7B-Base, Qwen3-4B-Base, Qwen3-8B-Base, Qwen3-14B-Base, ai-forever/mGPT, ai-forever/mGPT-13B, meta-llama/Llama-3.1-8B, Llama-3.1-70B, Llama-3.2-1B and Llama-3.2-3B. The word cells are aggregated from the token tables (derivation_source in tags.parquet).
Text. The RSC release writes most commas in item.text as U+201A (single low-9 quotation mark). These cells are scored on the text with that character read as a comma, which is what the toolkit's RSC() loads by default (RSC(standard_commas=True), replica-dev pull 564, "Read RSC's U+201A characters as commas by default", merged as 0527a1156). Every manifest records corpus digest bfd5ebd5. The cells supersede the rsc cells of pull request 8, which were scored on the released characters.
Start states. The Qwen3 Base models start from <|endoftext|> (--start-ids 151643), mGPT and mGPT-13B from <|endoftext|> (id 5), the Llama models from their BOS token (<|begin_of_text|>, id 128000).
Precision and agreement with the paper's values. Llama 3.1 70B ran in bfloat16 with device_map=auto over four A100-SXM4-80GB GPUs with torch 2.13.0 and transformers 4.57.6, the environment in which the paper's own cell was computed; its word and token values are bit-identical to the values the paper's analysis used (payload digests 9552472b and dd243750). This replaces the first version of these two cells on this pull request, computed on RTX PRO 6000 GPUs with torch 2.11.0, which differed from the paper's values by up to 0.45 bits. The other models ran in single precision, the driver's default. For the six models with a reference in the paper's analysis (mGPT, mGPT-13B, Llama 3.1 8B, Llama 3.1 70B, Llama 3.2 1B, Llama 3.2 3B) every word and token value equals the reference exactly; the Qwen3 models have no reference.
Code. The Llama 3.1 70B cells record code_commit 0527a1156 (main). The other 20 cells record 4391cfd3e, a commit on the pull 564 branch whose source digest (fc8baa87, source_sha256 in the manifests) equals the digest of main at 0527a1156, so every cell is reproducible from main.
Files. rsc.parquet, tokens/rsc.parquet, 22 manifests, tags.parquet, coverage.parquet, README.md (one rsc and one tokens_rsc config, and a sentence on the comma normalization), publish.json, readback.json, history.parquet. Base revision 7ef14c5c3 (main); verified on the stage before publishing.
Pull request 13 was merged on 2026-11-12 at 05:43 UTC with the first version of the two Llama 3.1 70B cells (RTX PRO 6000 GPUs, torch 2.11.0, up to 0.45 bits from the values the paper's analysis used). The two commits that reached this pull request's ref after the merge, 64c32bda810d and fc514c788736, carry the cells computed in the paper's environment and are not in main. Those cells are published as pull request 15 on top of main a5b12f094f8f.