{ "loader_kwargs": { "dtype": "bfloat16" }, "name": "KaLM-Embedding/KaLM-Reranker-V1-Small-R2", "revision": "68b74f09ff2ab5df9b9a9a48c812a81aaff8815e", "release_date": "2026-09-08", "languages": [ "afr-Latn", "ara-Arab", "aze-Latn", "bel-Cyrl", "bul-Cyrl", "ben-Beng", "cat-Latn", "ceb-Latn", "ces-Latn", "cym-Latn", "dan-Latn", "deu-Latn", "ell-Grek", "eng-Latn", "spa-Latn", "est-Latn", "eus-Latn", "fas-Arab", "fin-Latn", "fra-Latn", "glg-Latn", "guj-Gujr", "heb-Hebr", "hin-Deva", "hrv-Latn", "hat-Latn", "hun-Latn", "hye-Armn", "ind-Latn", "isl-Latn", "ita-Latn", "jpn-Jpan", "jav-Latn", "kat-Geor", "kaz-Cyrl", "khm-Khmr", "kan-Knda", "kor-Hang", "kir-Cyrl", "lao-Laoo", "lit-Latn", "lav-Latn", "mkd-Cyrl", "mal-Mlym", "mon-Cyrl", "mar-Deva", "msa-Latn", "mya-Mymr", "nep-Deva", "nld-Latn", "nor-Latn", "nob-Latn", "nno-Latn", "pan-Guru", "pol-Latn", "por-Latn", "que-Latn", "ron-Latn", "rus-Cyrl", "sin-Sinh", "slk-Latn", "slv-Latn", "swa-Latn", "tam-Taml", "tel-Telu", "tha-Thai", "tgl-Latn", "tur-Latn", "ukr-Cyrl", "urd-Arab", "vie-Latn", "yor-Latn", "zho-Hans" ], "n_parameters": 2115977456, "n_active_parameters_override": null, "n_embedding_parameters": 301989888, "memory_usage_mb": 4036.0, "max_tokens": 131072.0, "embed_dim": 1152, "license": "apache-2.0", "open_weights": true, "public_training_code": "https://github.com/HITsz-TMG/Lychee-Reranker", "public_training_data": "https://huggingface.co/datasets/KaLM-Embedding/KaLM-reranker-training-data", "framework": [ "PyTorch", "Transformers", "safetensors" ], "reference": "https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Small-R2", "similarity_fn_name": null, "use_instructions": true, "training_datasets": [ "MEDI2BGE", "THUCNews", "AdvertiseGen", "RefGPT", "webgpt_comparisons", "cCOVID-News", "LCSTS", "yahoo-answers", "lawzhidao", "FEVER", "CodeFeedback", "lima-chinese", "ccnews", "law-gpt", "PubMedQA", "HotpotQA", "retrieval_data_llm_infgrad", "SearchQA", "esci", "CMedQAv2-reranking", "csl", "T2Retrieval", "mmarco-chinese", "MSMARCOv2", "T2Reranking", "dureader-checklist", "PAQ", "NQ", "NFCorpus", "aya_dataset", "ArguAna", "DBPedia", "TRECCOVID", "MSMARCO", "arxiv_qa", "SciFact", "rag-dataset-12000", "UMETRIP-QA", "triviaqa", "CMRC 2018", "squad_v2", "QuoraRetrieval", "webqa", "DRCD", "MLDR", "GooAQ", "WebCPM", "FiQA2018", "CHEF", "Multi-CPR", "MrTidyRetrieval", "DuRetrieval", "Expertqa", "MIRACLRetrieval", "CQADupstackRetrieval", "OpenOrca" ], "adapted_from": "google/t5gemma-2-1b-1b", "superseded_by": null, "modalities": [ "text" ], "model_type": [ "cross-encoder" ], "citation": "@misc{zhao2026lycheererankerfastlateinteraction,\n title={Lychee-Reranker: Fast but Not Late Interaction for Compressed Document Reranking},\n author={Xinping Zhao and Jiaxin Xu and Ziqi Dai and Xin Zhang and Huiyao Chen and Shouzheng Huang and Xianhao Xiong and Danyu Tang and Xinshuo Hu and Guohong Fu and Meishan Zhang and Baotian Hu},\n year={2026},\n eprint={2606.22807},\n archivePrefix={arXiv},\n primaryClass={cs.CL},\n url={https://arxiv.org/abs/2606.22807},\n}", "contacts": null, "experiment_kwargs": null, "output_dtypes": "bfloat16", "extra_requirements_groups": null, "loader": "KaLMRerankerWrapper", "is_cross_encoder": true }