Commit History
-
Pin the base revision that was missing
Verified signatureAn adapter without a base pin resolves its checkpoint to whatever intfloat/multilingual-e5-small happens to be at load time. That is how the playground crash-looped once already, and embed-legal-en has carried this field all along while these two silently did not. The revision is not reconstructed from memory. The base repo's weights were uploaded once in 2023; every commit after 960310a7 (2023-07-12, the model_max_length fix) touches only the README, ONNX/OpenVINO exports and evaluation metadata. Checked file by file: config.json, model.safetensors, pytorch_model.bin, tokenizer.json, tokenizer_config.json, sentencepiece.bpe.model and special_tokens_map.json are byte-identical from 960310a7 through current main. So this sha names the same weights any training run since 2023 loaded, whichever commit it actually fetched.bf529e5
-
v1.1 (gov-indic-e4-ep6): +53% cross-lingual R@1, significant over base and v1.0; gu/or regress vs v1.0 (kept at tag v1.0)
Verified signature49411fc
-
Pin base checkpoint revision in adapter.json (self-resolves base offline)
Verified signaturee012172
-
embed-gov-indic v1 — cross-lingual government retrieval across 16 Indian languages (gov-indic-e3, +27.9% R@1, significant)
Verified signature26b5de5