Commit History
2 days ago
-
Add gemini-4-neon@p3 predictions (low thinking, batches of 50)
thewh1teagle +1: Claude Fable 5.2Unverified signatureCo-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>8bf223b
8 days ago
-
Add gemini-4-hydrogen@p3 predictions (low thinking, batches of 50)
thewh1teagle +1: Claude Fable 5.2Unverified signatureCo-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>0db2cec
11 days ago
-
Add claude-fable-5.2@p3 predictions (blind subagents, batches of 237)
thewh1teagle +1: ClaudeRun as seven Claude Code subagents on claude-fable-5-2, each given only row id and sentence: no labels, target indexes or categories. Each phonemized the whole sentence with p3 and the target words were taken by token index afterwards. Scores 0.7665 exact / 0.8579 1-WER / 0.9672 1-CER, second on exact match behind gemini-4-argon@p3 (0.7731). Every other column is byte-identical. Rows 940 and 1129 have gold target indexes past the sentence, so only in-range targets are filled. Reasoning effort was not recorded. Co-Authored-By: Claude <noreply@anthropic.com>b7ebbf8
-
Add gemini-4-argon@p3 predictions (low thinking, batches of 50)
thewh1teagleUnverified signature0cdf637
-
Keep the column named conikud (preview)
thewh1teagle +1: Claude Opus 5Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>34ae8e5 -
conikud: refresh predictions from the current model
thewh1teagle +1: Claude Opus 5Re-decoded all 1,653 rows. Exact 72.7 -> 73.9, 1-WER 83.9 -> 84.7, 1-CER 95.9 -> 96.1. Every other system's column is byte-identical; only this one moved. The model now also conditions on speaker and listener gender, which this benchmark does not exercise -- no row supplies either, so the numbers above are its unconditioned behaviour. Column renamed from "conikud (preview)" to "conikud". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>c79e519
-
Record a source URL in the meta row for the non-LLM systems
thewh1teagle +1: Claude Opus 5 (1M context)The leaderboard links each model name at its home page, and the meta row is already where per-system provenance lives, so the URL belongs there rather than hardcoded in the Space. Only systems that have a page get one; the LLM columns are identified by name alone. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CYLj1GjoMWDV5vcvDNdN9K3febaf6 -
Add conikud (preview) predictions; link the leaderboard
thewh1teagle +1: Claude Opus 5 (1M context)conikud is a local ONNX G2P rather than an LLM, so the column takes no prompt id. Run from an fp32 export on CUDA: the published int8 model has no native CUDA kernels for its integer matmuls and falls back across the device boundary, so fp32 is both faster on GPU and free of quantization rounding. Scores 0.8389 1-WER / 0.7272 exact, first on exact match across the systems in the file. Row 940 is empty as it is for renikud and phonikud, since that row's Label indexes token 5 of a 5-token sentence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CYLj1GjoMWDV5vcvDNdN9K851fbfe
-
Add Astra medium benchmark predictions and prompts; keep plans local
thewh1teagleUnverified signature045c17b
-
Add Gemini 3.7 Flash benchmark predictions
thewh1teagleUnverified signatureea6b31f
-
Align category names with the paper
thewh1teagle +1: Claude Opus 5Unverified signatureThe paper defines eleven categories of 150-151 sentences each, totalling 1,653. This repo had twelve, splitting homographs into "homographs" (101, both readings in one sentence) and "Regular Homographed" (50, readings split across sentence pairs). Merging them gives 151 rows and reproduces the paper's Gemini homograph score of 5.9 WER exactly, confirming they are the paper's single Homographs category. Also drops the -test suffix from ilspeech-v2-test. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>37ecd8d -
Record the prompt behind each LLM column
thewh1teagle +1: Claude Opus 5Unverified signatureLLM columns are now named model@id, where the id matches a section in PROMPTS.md. Columns sharing an id ran the same prompt and are comparable. - PROMPTS.md: p1 (gemini) and p2 (the write-only subagent prompt used for the two Claude columns), recorded verbatim as the models saw them - PROMPT.md removed; it duplicated whichever section is current - predictions.tsv columns renamed to carry their prompt id Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>1673e71 -
Add leaderboard, contribution docs, and three model columns
thewh1teagle +1: Claude Opus 5Unverified signature- data/predictions.tsv: add claude-opus-5-medium, claude-fable-5-medium, gemini-3.1-pro columns (1,653 rows each) - PROMPT.md: the phonemization prompt for LLM submissions, kept separate so it can be versioned independently of the contribution guide - CONTRIBUTING.md: how to run a model and submit a prediction column - space/: static HF Space leaderboard, reads gold.tsv and predictions.tsv from this repo at runtime so new columns appear without a code change Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>e6bee52
-
Upload colloquial_formal_informal_renikud_study_150.tar.gz
Verified signatured7407dc
-
Implement feature X to enhance user experience and fix bug Y in module Z
thewh1teagleUnverified signature842d811 -
init
thewh1teagleUnverified signaturea5bbc9e