Leave-k-out model bank for an EK-FAC linear-datamodeling-score (LDS) study of how attribution quality scales with training-set size N (this repo: N=4k 512-token chunks). GPT-2 fine-tuned on the 4k subset of...
Fonte do modelo
Descrição da fonte
Leave-k-out model bank for an EK-FAC linear-datamodeling-score (LDS) study of
how attribution quality scales with training-set size N (this repo: N=4k
512-token chunks). GPT-2 fine-tuned on the 4k subset of
EleutherAI/bergson-smollm2-lds-chunks (packed from
EleutherAI/SmolLM2-135M-10B).
Fontes
1 fonteVerificado 16 de jul.
Artefatos de modelo
1 artefatoTrechos de fonte
2 trechos--- datasets: - EleutherAI/bergson-smollm2-lds-chunks base_model: - gpt2 tags: - influence-functions - linear-datamodeling-score - ekfac --- # EleutherAI/bergson-smollm2-lds-4k Leave-k-out model bank for an EK-FAC linear-datamodeling-score (LDS) study of how attribution quality scales with training-set size N (this repo: N=4k 512-token chunks). GPT-2 fine-tuned on the `4k` subset of [`EleutherAI/bergson-smollm2-lds-chunks`](https://huggingface.co/datasets/EleutherAI/bergson-smollm2-lds-chunks) (packed from `EleutherAI/SmolLM2-135M-10B`). - `base/`: GPT-2 fine-tuned on the full N-chunk training set (no leave-out). - `subset_0/` … `subset_99/`: each retrains from scratch with a random 1%% of the training chunks held out, forming the LDS leave-k-out bank. Queries: the `query` config of the dataset repo (50 held-out chunks).
base/: GPT-2 fine-tuned on the full N-chunk training set (no leave-out).subset_0/ ... subset_99/: each retrains from scratch with a
random 1%% of the training chunks held out, forming the LDS leave-k-out bank.Queries: the query config of the dataset repo (50 held-out chunks).
Source context: 0 downloads · 0 likes · Repo EleutherAI/bergson-smollm2-lds-4k