ettin-reranker-68m-v1

This is a Cross Encoder model finetuned from jhu-clsp/ettin-encoder-68m on the cross-encoder/ettin-reranker-v1-data dataset using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.

See the release blogpost for details on the training recipe, evaluation results, and speed benchmarks against other public rerankers. The Evaluation section below also has the headline numbers.

Model Details

Model Description

• Model Type: Cross Encoder
• Base model: jhu-clsp/ettin-encoder-68m
• Maximum Sequence Length: 7999 tokens
• Number of Output Labels: 1 label
• Supported Modality: Text
• Training Dataset: cross-encoder/ettin-reranker-v1-data
• Language: en
• License: apache-2.0

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Cross Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Cross Encoders on Hugging Face

Full Model Architecture

CrossEncoder(
  (0): Transformer({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'architecture': 'ModernBertModel'})
  (1): Pooling({'embedding_dimension': 512, 'pooling_mode': 'cls', 'include_prompt': True})
  (2): Dense({'in_features': 512, 'out_features': 512, 'bias': False, 'activation_function': 'torch.nn.modules.activation.GELU', 'module_input_name': 'sentence_embedding', 'module_output_name': 'sentence_embedding'})
  (3): LayerNorm({'dimension': 512})
  (4): Dense({'in_features': 512, 'out_features': 1, 'bias': True, 'activation_function': 'torch.nn.modules.linear.Identity', 'module_input_name': 'sentence_embedding', 'module_output_name': 'scores'})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder(
    "cross-encoder/ettin-reranker-68m-v1",
    model_kwargs={"dtype": "bfloat16", "attn_implementation": "flash_attention_2"},  # Optional: pip install kernels
)

# Get scores for pairs of inputs
query = "Which planet is known as the Red Planet?"
passages = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet.",
]
scores = model.predict([(query, passage) for passage in passages])
print(scores)
# [ 6.375  11.5     7.625  10.4375]

# Or rank passages by relevance to a single query
ranked = model.rank(query, passages)
print(ranked)
# [{'corpus_id': 1, 'score': np.float32(11.5)}, ...]

Evaluation

MTEB(eng, v2) Retrieval

Each model in the ettin-reranker-v1 family was evaluated on the full MTEB(eng, v2) Retrieval benchmark (10 tasks, top-100 reranked) using MTEB's two-stage reranking flow, pairing each reranker with six embedding models that span the speed/quality spectrum. The dashed retriever-only line in each chart below is the headline number to beat. Anything below it means the reranker actively hurts the pipeline on average:

Mean NDCG@10 over the 6 embedder pairings, sorted by MTEB. The released ettin-reranker-v1 family is in bold, and the teacher mixedbread-ai/mxbai-rerank-large-v2 is underlined.

† Capped to max_seq_length=8192 (the 4B Qwen3-based rerankers don't fit on a single H100 80GB at native context). Native-context evaluation is likely higher.

See the release blogpost for the full analysis and per-model commentary.

Speed

All six released models were benchmarked against thirteen public rerankers on three hardware tiers, using sentence-transformers/natural-questions at max_length=512 with each model's best supported attention implementation. The full sweep over fp32+SDPA, bf16+SDPA, padded bf16+FA2, and unpadded bf16+FA2 (showing why the ettin-reranker-v1 family is faster than other ModernBERT-based rerankers) is in the release blogpost. This table shows the throughput in pairs per second on a NVIDIA H100 80GB, all in bfloat16:

Metrics

Cross Encoder Reranking

• Datasets: NanoMSMARCO_R100, NanoNFCorpus_R100, NanoNQ_R100, NanoFiQA2018_R100, NanoTouche2020_R100, NanoSciFact_R100, NanoHotpotQA_R100, NanoArguAna_R100, NanoFEVER_R100, NanoDBPedia_R100, NanoClimateFEVER_R100, NanoSCIDOCS_R100 and NanoQuoraRetrieval_R100
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

  {
      "at_k": 10,
      "always_rerank_positives": true
  }
  

Cross Encoder Nano BEIR

• Dataset: NanoBEIR_R100_mean
• Evaluated with CrossEncoderNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq",
          "fiqa2018",
          "touche2020",
          "scifact",
          "hotpotqa",
          "arguana",
          "fever",
          "dbpedia",
          "climatefever",
          "scidocs",
          "quoraretrieval"
      ],
      "dataset_id": "sentence-transformers/NanoBEIR-en",
      "rerank_k": 100,
      "at_k": 10,
      "always_rerank_positives": true
  }
  

[!NOTE] The release blogpost quotes a slightly higher NanoBEIR mean NDCG@10 of 0.6915 for this model, computed in fp32 rather than the bfloat16 used by the training-time evaluation above. Both numbers are valid.

Training Details

Training Dataset

ettin-reranker-v1-data

• Dataset: cross-encoder/ettin-reranker-v1-data
• Size: 143,393,475 training samples
• Columns: query, document, and label
• Approximate statistics based on the first 1000 samples:

  	query	document	label
  type	string	string	float
  details	min: 26 charactersmean: 55.52 charactersmax: 249 characters	min: 63 charactersmean: 659.91 charactersmax: 3975 characters	min: -2.94mean: 8.51max: 13.88

• Samples:

  query	document	label
  Help me with my Reborn performance	I was reading the comment section for Dotacinema's world of dota video, and a bunch of people were complaining how there were a lot of bugs and some talked about PERFORMANCE ISSUES. But there were also people saying that reborn has actually IMPROVED their gameplay?I am one of those people who is running into performance issues and would desperately like to know how some are getting BETTER performance while others like me are getting worse. I'm not complaining about bugs, I'm complaing about framerate, I use to get 60 fps solid in source 1 but I now have 40 or at worst 30 fps in source 2.I have an i3 processor/gtx560ti/16gb RAMi dont think it's a potato pc, so I dont know what's happening, I cleaned my computer recently so dust isnt affecting anything in anyway.So if you gained or had IMPROVED performance in source 2 please list the settings you are enabling, so I can see where I am at fault. (v sync is off btw)TLDR: Have bad performance now from source 2, if you have good p...	9.5
  Really wanna try out the game and expansion, ~$60 is hefty. Likelihood of sales?	As per title, steam sells the game and its expansions for $60 total. Heavy price to drop. Are there sales on any other website? This game looks fantastic to immerse in otherwise and I'm pleased that this subreddit has at least some attention to help out new folks!	9.25
  Your Avatar. [MGSV Spoilers]	Was anyone else suprised he actually replaces the snake model in some cutscenes. I've only tried the first Quiet cutscenes, i was just amazed I haven't seen anybody else say this yet.Sorry if repost.	5.25

• Loss: MSELoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity"
  }
  

Evaluation Dataset

ettin-reranker-v1-data

• Dataset: cross-encoder/ettin-reranker-v1-data
• Size: 5,000 evaluation samples
• Columns: query, document, and label
• Approximate statistics based on the first 1000 samples:

  	query	document	label
  type	string	string	float
  details	min: 14 charactersmean: 52.62 charactersmax: 168 characters	min: 11 charactersmean: 50.12 charactersmax: 184 characters	min: 4.44mean: 13.49max: 18.62

• Samples:

  query	document	label
  Why do we need binomial distribution?	Why is the binomial distribution important?	11.375
  I already have Windows 10, can I delete Windows.old?	After resetting windows 10, can I safely delete the "old windows" folder?	10.875
  How can guys last longer during sex?	How do men last longer in bed?	10.8125

• Loss: MSELoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 16
• num_train_epochs: 1
• learning_rate: 3e-05
• warmup_steps: 0.03
• bf16: True
• per_device_eval_batch_size: 16
• load_best_model_at_end: True
• seed: 12

All Hyperparameters

• per_device_train_batch_size: 16
• num_train_epochs: 1
• max_steps: -1
• learning_rate: 3e-05
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 0.03
• optim: adamw_torch
• optim_args: None
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 1
• average_tokens_across_devices: True
• max_grad_norm: 1.0
• label_smoothing_factor: 0.0
• bf16: True
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: True
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: False
• project: huggingface
• trackio_space_id: None
• trackio_bucket_id: None
• trackio_static_space_id: None
• per_device_eval_batch_size: 16
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: False
• hub_private_repo: None
• hub_model_id: None
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: True
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 12
• data_seed: None
• use_cpu: False
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• dataloader_drop_last: True
• dataloader_num_workers: 0
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_static_graph: None
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: []
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• warmup_ratio: None
• local_rank: -1
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Training Time

• Training: 11.2 hours
• Evaluation: 9.2 minutes
• Total: 11.3 hours

Framework Versions

• Python: 3.11.15
• Sentence Transformers: 5.4.1
• Transformers: 5.7.0
• PyTorch: 2.7.0+cu126
• Accelerate: 1.13.0
• Datasets: 4.8.5
• Tokenizers: 0.22.2

Citation

BibTeX

Ettin Reranker Blogpost

@misc{aarsen2026ettin-reranker,
    title = "Introducing the Ettin Reranker Family",
    author = "Aarsen, Tom",
    year = "2026",
    publisher = "Hugging Face",
    url = "https://huggingface.co/blog/ettin-reranker",
}

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}