main

GGitHub[fix] Pass 1-based ranks to the custom rank_discount_fn in PListMLELambdaWeight (#4204)

5336c6a9创建于 8 小时前历史提交

文件	最后提交记录	最后更新时间
cross_encoder	[fix] Pass 1-based ranks to the custom rank_discount_fn in PListMLELambdaWeight (#4204) * [fix] Pass 1-based ranks to the custom rank_discount_fn in PListMLELambdaWeight PListMLELambdaWeight documents rank_discount_fn as receiving rank positions (e.g. 1.0 / torch.log1p(ranks) with ranks [1, 2, 3, ...]), but it was called with the boolean validity mask. The documented discount then gave every valid position the same weight and padded positions an infinite one, which normalizes to NaN and zeroes the loss of every list shorter than the longest in the batch. Pass the 1-based ranks instead and zero the weights of padded positions. * Clarify custom rank discount docs and simplify regression test * Preserve mask callbacks and correct custom rank discount examples --------- Co-authored-by: Tom Aarsen <Cubiegamedev@gmail.com>	8 小时前
multi_vector_encoder	🚨 Normalize hierarchical token pooling outputs by default (#4008) * Normalize hierarchical token pooling outputs by default * Normalize unpooled token batches without per-document overhead * Compute single-cluster token means without hierarchical clustering * Use a compatible zero-centroid fixture instead of changing pooling	3 天前
sentence_transformer	[examples] Add PEP 723 script headers to the STS training examples (#4072) Adds a PEP 723 header to the STS training examples for Sentence Transformer, Cross Encoder and Sparse Encoder models, so they run with `uv run` or on Hugging Face Jobs straight from their URL. The STS README gets a short section with the Jobs command. Claude-Session: https://claude.ai/code/session_01VDeRk1QQEvGKZzwdb2VHsS Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	12 天前
sparse_encoder	[examples] Add PEP 723 script headers to the STS training examples (#4072) Adds a PEP 723 header to the STS training examples for Sentence Transformer, Cross Encoder and Sparse Encoder models, so they run with `uv run` or on Hugging Face Jobs straight from their URL. The STS README gets a short section with the Jobs command. Claude-Session: https://claude.ai/code/session_01VDeRk1QQEvGKZzwdb2VHsS Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	12 天前