tomaarsen/reranker-camembertv2-base-fr-lambda

CrossEncoder based on almanach/camembertv2-base

This is a Cross Encoder model finetuned from almanach/camembertv2-base 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: almanach/camembertv2-base
• Maximum Sequence Length: 1024 tokens
• Number of Output Labels: 1 label

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("tomaarsen/reranker-camembertv2-base-fr-lambda")
# Get scores for pairs of texts
pairs = [
    ['How many calories in an egg', 'There are on average between 55 and 80 calories in an egg depending on its size.'],
    ['How many calories in an egg', 'Egg whites are very low in calories, have no fat, no cholesterol, and are loaded with protein.'],
    ['How many calories in an egg', 'Most of the calories in an egg come from the yellow yolk in the center.'],
]
scores = model.predict(pairs)
print(scores.shape)
# (3,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'How many calories in an egg',
    [
        'There are on average between 55 and 80 calories in an egg depending on its size.',
        'Egg whites are very low in calories, have no fat, no cholesterol, and are loaded with protein.',
        'Most of the calories in an egg come from the yellow yolk in the center.',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Evaluation

Metrics

Cross Encoder Reranking

• Dataset: swim_ir_dev
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

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

Metric	Value
map	0.6059 (+0.1333)
mrr@10	0.6052 (+0.1371)
ndcg@10	0.6217 (+0.1206)

Training Details

Training Dataset

Unnamed Dataset

• Size: 100,000 training samples
• Columns: query, docs, and labels
• Approximate statistics based on the first 1000 samples:

  	query	docs	labels
  type	string	list	list
  details	min: 0 characters mean: 37.74 characters max: 157 characters	size: 6 elements	size: 6 elements

• Samples:

  query	docs	labels
  	['L'ambitus exigé par le rôle-titre est plus problématique : la plus haute note est un "si" aigu, ce qui n'est pas anormal pour une soprano ou une mezzo-soprano, alors que la plus basse est un "sol" bémol grave dans le registre alto (et normalement au-dessous du registre d'une mezzo-soprano "standard"). Compte tenu d'une telle tessiture, qui ressemble à celle de nombreux rôles de mezzo comme Carmen et Amneris, on pourrait croire qu'un soprano aigu n'est pas essentiel à la pièce, mais c'est bien le contraire ; la plupart des sopranos graves qui ont abordé ce rôle ont imposé un tel effort à leur voix tout au long de l'opéra, qu'elles se retrouvaient épuisées au moment de la scène finale (la partie la plus éprouvante pour le rôle-titre). Ce rôle est l'exemple classique de la différence qui existe entre tessiture et ambitus : tandis que des mezzos peuvent exécuter une note aigüe (comme dans "Carmen"), ou même soutenir temporairement une tessiture tendue, il est impossible pour un...	[1, 0, 0, 0, 0, ...]
  	["Les saisons 2 à 6 sont produites par Télé-Vision V Inc., filiale de Groupe Télé-Vision Inc. Lors de la saison d'hiver 2006, l'émission était animée par Isabelle Maréchal et Virginie Coossa. Pour les saisons 3,4,5 Marie Plourde a remplacé Isabelle Maréchal, alors que Virginie Coossa est demeurée coanimatrice. Lors des saisons 5 et 6, Kim Rusk, la gagnante de la saison 3, était la coanimatrice. La saison 6 sera animée par Pierre-Yves Lord.", ',{"type": "ExternalData", "service":"geoshape","ids": "Q40","properties": {"fill":"#FF0000","stroke-width":0,"description": "Autriche"}}]', '! scope=col width="10%"	Pages ! scope=col width="25%"
  	['En 1963, Bernard et Françoise Moitessier quittent le port de Marseille, pour un voyage de noces. Ils prennent le détroit de Gibraltar et se dirigent vers les îles Canaries où il retrouve Pierre Deshumeurs, le compagnon du "Snark". Les enfants de Françoise les rejoignent le temps des vacances scolaires. Les Moitessier poursuivent ensuite vers les Antilles, puis le canal de Panama, avant de s'arrêter longuement dans l'archipel des Galápagos, où certaines îles reculées de toutes civilisations accueillent une faune et une flore exceptionnelles qui retiennent l'attention du couple. Ils rejoignent ensuite la Polynésie française où ils restent plusieurs mois.', ',{"type": "ExternalData", "service":"geoshape","ids": "Q40","properties": {"fill":"#FF0000","stroke-width":0,"description": "Autriche"}}]', '! scope=col width="10%"	Pages ! scope=col width="25%"

• Loss: LambdaLoss with these parameters:

  {
      "weighting_scheme": "sentence_transformers.cross_encoder.losses.LambdaLoss.NDCGLoss2PPScheme",
      "k": null,
      "sigma": 1.0,
      "eps": 1e-10,
      "reduction_log": "binary",
      "activation_fct": "torch.nn.modules.linear.Identity",
      "mini_batch_size": 8
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True
• 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: 16
• per_device_eval_batch_size: 16
• 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: True
• 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}
• 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
• dispatch_batches: None
• split_batches: 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

Training Logs

Epoch	Step	Training Loss	swim_ir_dev_ndcg@10
-1	-1	-	0.0784 (-0.4228)
0.0002	1	2.0475	-
0.016	100	2.065	-
0.032	200	1.9662	-
0.048	300	0.9965	-
0.064	400	0.7667	-
0.08	500	0.6547	0.5961 (+0.0950)
0.096	600	0.5899	-
0.112	700	0.5331	-
0.128	800	0.4637	-
0.144	900	0.4826	-
0.16	1000	0.4249	0.6012 (+0.1000)
0.176	1100	0.4271	-
0.192	1200	0.4071	-
0.208	1300	0.3594	-
0.224	1400	0.401	-
0.24	1500	0.4171	0.5900 (+0.0888)
0.256	1600	0.3728	-
0.272	1700	0.3242	-
0.288	1800	0.3665	-
0.304	1900	0.3367	-
0.32	2000	0.3259	0.6134 (+0.1122)
0.336	2100	0.381	-
0.352	2200	0.3289	-
0.368	2300	0.3234	-
0.384	2400	0.3794	-
0.4	2500	0.3322	0.6070 (+0.1058)
0.416	2600	0.3139	-
0.432	2700	0.3427	-
0.448	2800	0.3162	-
0.464	2900	0.2899	-
0.48	3000	0.3571	0.6166 (+0.1155)
0.496	3100	0.3312	-
0.512	3200	0.3082	-
0.528	3300	0.2839	-
0.544	3400	0.3649	-
0.56	3500	0.325	0.6108 (+0.1097)
0.576	3600	0.3042	-
0.592	3700	0.2785	-
0.608	3800	0.3095	-
0.624	3900	0.3053	-
0.64	4000	0.293	0.6131 (+0.1119)
0.656	4100	0.2987	-
0.672	4200	0.2675	-
0.688	4300	0.2977	-
0.704	4400	0.2881	-
0.72	4500	0.2862	0.6186 (+0.1174)
0.736	4600	0.2996	-
0.752	4700	0.2724	-
0.768	4800	0.2442	-
0.784	4900	0.2923	-
0.8	5000	0.2691	0.6217 (+0.1206)
0.816	5100	0.3042	-
0.832	5200	0.2654	-
0.848	5300	0.3059	-
0.864	5400	0.2571	-
0.88	5500	0.2741	0.6174 (+0.1162)
0.896	5600	0.3009	-
0.912	5700	0.2669	-
0.928	5800	0.2272	-
0.944	5900	0.2673	-
0.96	6000	0.2674	0.6194 (+0.1182)
0.976	6100	0.2551	-
0.992	6200	0.2981	-
-1	-1	-	0.6217 (+0.1206)

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.680 kWh
• Carbon Emitted: 0.264 kg of CO2
• Hours Used: 1.921 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.11.6
• Sentence Transformers: 3.5.0.dev0
• Transformers: 4.49.0
• PyTorch: 2.6.0+cu124
• Accelerate: 1.4.0
• Datasets: 3.3.2
• Tokenizers: 0.21.0

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",
}

LambdaLoss

@inproceedings{wang2018lambdaloss,
  title={The lambdaloss framework for ranking metric optimization},
  author={Wang, Xuanhui and Li, Cheng and Golbandi, Nadav and Bendersky, Michael and Najork, Marc},
  booktitle={Proceedings of the 27th ACM international conference on information and knowledge management},
  pages={1313--1322},
  year={2018}
}