bansalaman18/reranker-msmarco-v1.1-ettin-encoder-68m-bce

CrossEncoder based on jhu-clsp/ettin-encoder-68m

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

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
• Training Dataset:
  • ms_marco
• Language: en

Model Sources

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

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("bansalaman18/reranker-msmarco-v1.1-ettin-encoder-68m-bce")
# Get scores for pairs of texts
pairs = [
    ['how to put word count on word', 'To insert a word count into a Word 2013 document, place the cursor where you would like the word count to appear (say in the Header or Footer) and then: 1  click the Insert tab. 2  click the Quick Parts icon (towards the right hand end of the toolbar). 3  on the drop down that appears, select Field...'],
    ['what is the difference between discipleship and evangelism', 'Discipleship, on the other hand, meant helping someone who was already a believer walk out the life of faith. The word “discipleship” brought to my mind a small group Bible study, a conversation across the table with another woman, or an accountability group. And I knew which one I preferred. As a result, the discipleship I offered others contained a lot of good information but lacked the transforming power that can only come from the gospel. (I was also, simply, a coward.). I am beginning to see that evangelism and discipleship are not all that different.'],
    ['what metal is a trophy made from', 'The trophy stands 36.5 centimetres (14.4 inches) tall and is made  of 5 kg (11 lb) of 18 carat (75%) gold with a base (13 centimetres  [5.1 inches] in diameter) containing t … wo layers of malachite. Making the world better, one answer at a time. Trophies can be made out of anything you want. however, aluminum is a very reliable and trustworthy metal and it.......... oh crap.......... i have to do a poo...'],
    ['how do you define what a cult is?', 'The term cult has been misused. The word cult comes from the French cult which is from the Latin word cultus (care/adoration) and Latin Colere (to cultivate.) So, we can plant seeds of good or bad. You can have political cults such as sit ins during the Vietnam War. A good cult could be a religious one, yet some Christians will consider Jehovah Witness a cult and have labeled them as preying on the weak. When someone labels such a thing it is usually because of the lack of understanding. Good cults are usually a small group of people that can have a cult in most anything.'],
    ['where is silchar', 'Silchar (/ˈsɪlˌʧə/ or /ˈʃɪlˌʧə/) (Bengali: শিলচর Shilchor) shilchôr is the headquarters Of cachar district in the state Of assam In. India it is 343 (kilometres 213) mi south east Of. Guwahati it is the-second largest city of the state in terms of population and municipal. area 1 The Bhubaneshwar temple is about 50 km from Silchar and is on the top the Bhuvan hill. 2  This is a place of pilgrimage and during the festival of Shivaratri, thousand of Shivayats march towards the hilltop to worship Lord Shiva.'],
]
scores = model.predict(pairs)
print(scores.shape)
# (5,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'how to put word count on word',
    [
        'To insert a word count into a Word 2013 document, place the cursor where you would like the word count to appear (say in the Header or Footer) and then: 1  click the Insert tab. 2  click the Quick Parts icon (towards the right hand end of the toolbar). 3  on the drop down that appears, select Field...',
        'Discipleship, on the other hand, meant helping someone who was already a believer walk out the life of faith. The word “discipleship” brought to my mind a small group Bible study, a conversation across the table with another woman, or an accountability group. And I knew which one I preferred. As a result, the discipleship I offered others contained a lot of good information but lacked the transforming power that can only come from the gospel. (I was also, simply, a coward.). I am beginning to see that evangelism and discipleship are not all that different.',
        'The trophy stands 36.5 centimetres (14.4 inches) tall and is made  of 5 kg (11 lb) of 18 carat (75%) gold with a base (13 centimetres  [5.1 inches] in diameter) containing t … wo layers of malachite. Making the world better, one answer at a time. Trophies can be made out of anything you want. however, aluminum is a very reliable and trustworthy metal and it.......... oh crap.......... i have to do a poo...',
        'The term cult has been misused. The word cult comes from the French cult which is from the Latin word cultus (care/adoration) and Latin Colere (to cultivate.) So, we can plant seeds of good or bad. You can have political cults such as sit ins during the Vietnam War. A good cult could be a religious one, yet some Christians will consider Jehovah Witness a cult and have labeled them as preying on the weak. When someone labels such a thing it is usually because of the lack of understanding. Good cults are usually a small group of people that can have a cult in most anything.',
        'Silchar (/ˈsɪlˌʧə/ or /ˈʃɪlˌʧə/) (Bengali: শিলচর Shilchor) shilchôr is the headquarters Of cachar district in the state Of assam In. India it is 343 (kilometres 213) mi south east Of. Guwahati it is the-second largest city of the state in terms of population and municipal. area 1 The Bhubaneshwar temple is about 50 km from Silchar and is on the top the Bhuvan hill. 2  This is a place of pilgrimage and during the festival of Shivaratri, thousand of Shivayats march towards the hilltop to worship Lord Shiva.',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Evaluation

Metrics

Cross Encoder Reranking

• Datasets: NanoMSMARCO_R100, NanoNFCorpus_R100 and NanoNQ_R100
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

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

Metric	NanoMSMARCO_R100	NanoNFCorpus_R100	NanoNQ_R100
map	0.4957 (+0.0061)	0.3315 (+0.0706)	0.4563 (+0.0367)
mrr@10	0.4842 (+0.0067)	0.4747 (-0.0251)	0.4704 (+0.0437)
ndcg@10	0.5462 (+0.0058)	0.3291 (+0.0040)	0.5310 (+0.0304)

Cross Encoder Nano BEIR

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

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "rerank_k": 100,
      "at_k": 10,
      "always_rerank_positives": true
  }
  

Metric	Value
map	0.4279 (+0.0378)
mrr@10	0.4764 (+0.0084)
ndcg@10	0.4688 (+0.0134)

Training Details

Training Dataset

ms_marco

• Dataset: ms_marco at a47ee7a
• Size: 654,438 training samples
• Columns: query, response, and label
• Approximate statistics based on the first 1000 samples:

  	query	response	label
  type	string	string	float
  details	min: 11 characters mean: 34.1 characters max: 95 characters	min: 62 characters mean: 417.26 characters max: 939 characters	min: 0.0 mean: 0.13 max: 1.0

• Samples:

  query	response	label
  what is a polyhedron definition	A polyhedron is said to be convex if its surface (comprising its faces, edges and vertices) does not intersect itself and the line segment joining any two points of the polyhedron is contained in the interior or surface. A polyhedron is a 3-dimensional example of the more general polytope in any number of dimensions. Polyhedra with congruent regular faces of six or more sides are all non-convex, because the vertex of three regular hexagons defines a plane. The total number of convex polyhedra with equal regular faces is thus ten, comprising the five Platonic solids and the five non-uniform deltahedra.	0.0
  what can you carry in hand luggage on easyjet	Each passenger who pays for a hold bag can take up to 20kg of luggage. This weight allowance applies to the passenger rather than to the bag so purchasing extra bags is possible but will not increase the weight allowance.	0.0
  what is dynamic segmentation in gis	The result of the dynamic segmentation process is a dynamic feature class known as a route event source. A route event source can serve as the data source of a feature layer in ArcMap. For the most part, a dynamic feature layer behaves like any other feature layer. Event locating errors. The dynamic segmentation process creates a shape for each row in the input route event table. In some cases, however, the shape of the event feature might be empty. This happens when there is a reason that the event can't be properly located.	0.0

• Loss: BinaryCrossEntropyLoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity",
      "pos_weight": null
  }
  

Evaluation Dataset

ms_marco

• Dataset: ms_marco at a47ee7a
• Size: 1,000 evaluation samples
• Columns: query, response, and label
• Approximate statistics based on the first 1000 samples:

  	query	response	label
  type	string	string	float
  details	min: 10 characters mean: 33.73 characters max: 117 characters	min: 57 characters mean: 412.4 characters max: 918 characters	min: 0.0 mean: 0.13 max: 1.0

• Samples:

  query	response	label
  how to put word count on word	To insert a word count into a Word 2013 document, place the cursor where you would like the word count to appear (say in the Header or Footer) and then: 1 click the Insert tab. 2 click the Quick Parts icon (towards the right hand end of the toolbar). 3 on the drop down that appears, select Field...	0.0
  what is the difference between discipleship and evangelism	Discipleship, on the other hand, meant helping someone who was already a believer walk out the life of faith. The word “discipleship” brought to my mind a small group Bible study, a conversation across the table with another woman, or an accountability group. And I knew which one I preferred. As a result, the discipleship I offered others contained a lot of good information but lacked the transforming power that can only come from the gospel. (I was also, simply, a coward.). I am beginning to see that evangelism and discipleship are not all that different.	0.0
  what metal is a trophy made from	The trophy stands 36.5 centimetres (14.4 inches) tall and is made of 5 kg (11 lb) of 18 carat (75%) gold with a base (13 centimetres [5.1 inches] in diameter) containing t … wo layers of malachite. Making the world better, one answer at a time. Trophies can be made out of anything you want. however, aluminum is a very reliable and trustworthy metal and it.......... oh crap.......... i have to do a poo...	1.0

• Loss: BinaryCrossEntropyLoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity",
      "pos_weight": null
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True
• remove_unused_columns: False
• load_best_model_at_end: True

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 2e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 12
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: False
• label_names: None
• load_best_model_at_end: True
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• tp_size: 0
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	NanoMSMARCO_R100_ndcg@10	NanoNFCorpus_R100_ndcg@10	NanoNQ_R100_ndcg@10	NanoBEIR_R100_mean_ndcg@10
-1	-1	-	-	0.0442 (-0.4962)	0.2555 (-0.0695)	0.0464 (-0.4542)	0.1154 (-0.3400)
0.0002	1	2.1556	-	-	-	-	-
0.0196	100	0.84	0.3976	0.0474 (-0.4930)	0.2826 (-0.0424)	0.0395 (-0.4612)	0.1232 (-0.3322)
0.0391	200	0.402	0.3895	0.0508 (-0.4896)	0.2367 (-0.0883)	0.0673 (-0.4334)	0.1183 (-0.3371)
0.0587	300	0.3988	0.3854	0.0763 (-0.4642)	0.2175 (-0.1076)	0.1093 (-0.3913)	0.1344 (-0.3210)
0.0782	400	0.3899	0.3968	0.1461 (-0.3943)	0.2087 (-0.1163)	0.1574 (-0.3433)	0.1707 (-0.2846)
0.0978	500	0.3898	0.3708	0.2417 (-0.2987)	0.2259 (-0.0991)	0.2326 (-0.2680)	0.2334 (-0.2219)
0.1173	600	0.3783	0.3681	0.2933 (-0.2471)	0.2769 (-0.0481)	0.3659 (-0.1347)	0.3120 (-0.1433)
0.1369	700	0.3718	0.3610	0.3654 (-0.1751)	0.2852 (-0.0398)	0.4239 (-0.0768)	0.3582 (-0.0972)
0.1565	800	0.3643	0.3657	0.4721 (-0.0683)	0.2658 (-0.0593)	0.4801 (-0.0206)	0.4060 (-0.0494)
0.1760	900	0.3639	0.3637	0.4291 (-0.1113)	0.2635 (-0.0615)	0.4308 (-0.0699)	0.3745 (-0.0809)
0.1956	1000	0.3686	0.3518	0.4859 (-0.0545)	0.3071 (-0.0179)	0.5328 (+0.0322)	0.4419 (-0.0134)
0.2151	1100	0.3585	0.3529	0.4581 (-0.0823)	0.2611 (-0.0640)	0.5113 (+0.0107)	0.4102 (-0.0452)
0.2347	1200	0.3624	0.3522	0.4918 (-0.0486)	0.3244 (-0.0006)	0.4936 (-0.0071)	0.4366 (-0.0188)
0.2543	1300	0.3684	0.3607	0.3933 (-0.1471)	0.2893 (-0.0358)	0.4810 (-0.0196)	0.3879 (-0.0675)
0.2738	1400	0.3651	0.3437	0.4999 (-0.0406)	0.3170 (-0.0081)	0.5290 (+0.0283)	0.4486 (-0.0068)
0.2934	1500	0.3568	0.3466	0.5099 (-0.0306)	0.3312 (+0.0062)	0.4966 (-0.0040)	0.4459 (-0.0095)
0.3129	1600	0.36	0.3437	0.4762 (-0.0642)	0.3488 (+0.0237)	0.4963 (-0.0043)	0.4404 (-0.0149)
0.3325	1700	0.3557	0.3451	0.4572 (-0.0833)	0.3046 (-0.0204)	0.5173 (+0.0167)	0.4264 (-0.0290)
0.3520	1800	0.3529	0.3406	0.4707 (-0.0697)	0.3180 (-0.0071)	0.4842 (-0.0165)	0.4243 (-0.0311)
0.3716	1900	0.3514	0.3367	0.4901 (-0.0503)	0.2742 (-0.0509)	0.5322 (+0.0316)	0.4321 (-0.0232)
0.3912	2000	0.3499	0.3408	0.4859 (-0.0545)	0.2814 (-0.0437)	0.5011 (+0.0004)	0.4228 (-0.0326)
0.4107	2100	0.3595	0.3393	0.4821 (-0.0583)	0.3004 (-0.0246)	0.5499 (+0.0493)	0.4441 (-0.0112)
0.4303	2200	0.356	0.3442	0.4939 (-0.0465)	0.3279 (+0.0029)	0.5454 (+0.0448)	0.4557 (+0.0004)
0.4498	2300	0.3396	0.3351	0.5252 (-0.0152)	0.3024 (-0.0226)	0.5271 (+0.0264)	0.4516 (-0.0038)
0.4694	2400	0.3644	0.3396	0.5307 (-0.0098)	0.3204 (-0.0046)	0.5101 (+0.0094)	0.4537 (-0.0017)
0.4889	2500	0.3508	0.3371	0.5003 (-0.0402)	0.3006 (-0.0245)	0.5404 (+0.0398)	0.4471 (-0.0083)
0.5085	2600	0.3525	0.3396	0.5146 (-0.0258)	0.3001 (-0.0249)	0.5525 (+0.0518)	0.4557 (+0.0004)
0.5281	2700	0.3348	0.3393	0.4800 (-0.0604)	0.2778 (-0.0472)	0.5416 (+0.0410)	0.4332 (-0.0222)
0.5476	2800	0.3448	0.3458	0.5176 (-0.0229)	0.2905 (-0.0345)	0.5243 (+0.0236)	0.4441 (-0.0113)
0.5672	2900	0.3508	0.3379	0.4738 (-0.0667)	0.2924 (-0.0326)	0.5395 (+0.0389)	0.4352 (-0.0201)
0.5867	3000	0.3401	0.3404	0.5246 (-0.0158)	0.2930 (-0.0321)	0.5337 (+0.0330)	0.4504 (-0.0049)
0.6063	3100	0.3508	0.3383	0.5004 (-0.0400)	0.2890 (-0.0360)	0.5321 (+0.0314)	0.4405 (-0.0149)
0.6259	3200	0.3509	0.3364	0.5097 (-0.0308)	0.3321 (+0.0071)	0.5502 (+0.0496)	0.4640 (+0.0086)
0.6454	3300	0.3501	0.3369	0.5172 (-0.0232)	0.3084 (-0.0167)	0.5644 (+0.0637)	0.4633 (+0.0080)
0.6650	3400	0.3417	0.3336	0.4947 (-0.0457)	0.3133 (-0.0117)	0.5404 (+0.0397)	0.4495 (-0.0059)
0.6845	3500	0.3487	0.3335	0.4994 (-0.0410)	0.3328 (+0.0078)	0.5351 (+0.0344)	0.4558 (+0.0004)
0.7041	3600	0.3507	0.3377	0.5103 (-0.0301)	0.3111 (-0.0139)	0.5030 (+0.0023)	0.4415 (-0.0139)
0.7236	3700	0.34	0.3382	0.5254 (-0.0150)	0.3320 (+0.0070)	0.5154 (+0.0148)	0.4576 (+0.0023)
0.7432	3800	0.3392	0.3361	0.4892 (-0.0512)	0.3261 (+0.0011)	0.5268 (+0.0262)	0.4474 (-0.0080)
0.7628	3900	0.3511	0.3349	0.5129 (-0.0276)	0.3307 (+0.0057)	0.5180 (+0.0173)	0.4539 (-0.0015)
0.7823	4000	0.3508	0.3368	0.5462 (+0.0058)	0.3291 (+0.0040)	0.5310 (+0.0304)	0.4688 (+0.0134)
0.8019	4100	0.3439	0.3348	0.5409 (+0.0005)	0.3307 (+0.0056)	0.5312 (+0.0306)	0.4676 (+0.0122)
0.8214	4200	0.3487	0.3340	0.5324 (-0.0080)	0.3284 (+0.0034)	0.5191 (+0.0185)	0.4600 (+0.0046)
0.8410	4300	0.341	0.3351	0.5277 (-0.0127)	0.3276 (+0.0025)	0.5050 (+0.0044)	0.4535 (-0.0019)
0.8606	4400	0.3293	0.3344	0.5150 (-0.0254)	0.3355 (+0.0105)	0.5009 (+0.0002)	0.4505 (-0.0049)
0.8801	4500	0.3525	0.3346	0.5224 (-0.0180)	0.3239 (-0.0012)	0.5127 (+0.0121)	0.4530 (-0.0024)
0.8997	4600	0.3421	0.3331	0.5312 (-0.0092)	0.3376 (+0.0126)	0.5234 (+0.0228)	0.4641 (+0.0087)
0.9192	4700	0.3442	0.3336	0.5227 (-0.0177)	0.3330 (+0.0080)	0.5053 (+0.0047)	0.4537 (-0.0017)
0.9388	4800	0.3361	0.3328	0.5166 (-0.0238)	0.3378 (+0.0128)	0.5194 (+0.0187)	0.4579 (+0.0026)
0.9583	4900	0.3377	0.3335	0.5298 (-0.0106)	0.3300 (+0.0049)	0.5248 (+0.0241)	0.4615 (+0.0061)
0.9779	5000	0.3455	0.3332	0.5298 (-0.0106)	0.3305 (+0.0055)	0.5182 (+0.0176)	0.4595 (+0.0041)
0.9975	5100	0.3422	0.3333	0.5298 (-0.0106)	0.3303 (+0.0053)	0.5240 (+0.0234)	0.4614 (+0.0060)
-1	-1	-	-	0.5462 (+0.0058)	0.3291 (+0.0040)	0.5310 (+0.0304)	0.4688 (+0.0134)

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.13
• Sentence Transformers: 5.0.0
• Transformers: 4.51.0
• PyTorch: 2.9.1+cu126
• Accelerate: 1.8.1
• Datasets: 3.6.0
• Tokenizers: 0.21.4-dev.0

Citation

BibTeX

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",
}