You can deploy this model and start earning money today!

tomaarsen/
splade-distilbert-base-uncased-quora-duplicates

Name: splade-distilbert-base-uncased-quora-duplicates
Availability: InStock

Views

No views yet

splade-distilbert-base-uncased trained on Quora Duplicates Questions

This is a SPLADE Sparse Encoder model finetuned from distilbert/distilbert-base-uncased on the quora-duplicates dataset using the sentence-transformers library. It maps sentences & paragraphs to a 30522-dimensional sparse vector space and can be used for semantic search and sparse retrieval.

Model Details

Model Description

Model Type: SPLADE Sparse Encoder
Base model: distilbert/distilbert-base-uncased
Maximum Sequence Length: 256 tokens
Output Dimensionality: 30522 dimensions
Similarity Function: Dot Product
Training Dataset:
- quora-duplicates
Language: en
License: apache-2.0

Model Sources

Documentation: Sentence Transformers Documentation
Documentation: Sparse Encoder Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sparse Encoders on Hugging Face

Full Model Architecture

SparseEncoder(
  (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: DistilBertForMaskedLM 
  (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
)

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.

python

from sentence_transformers import SparseEncoder

# Download from the 🤗 Hub
model = SparseEncoder("tomaarsen/splade-distilbert-base-uncased-quora-duplicates")
# Run inference
sentences = [
    'What accomplishments did Hillary Clinton achieve during her time as Secretary of State?',
    "What are Hillary Clinton's most recognized accomplishments while Secretary of State?",
    'What are Hillary Clinton’s qualifications to be President?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 30522]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[ 83.9635,  60.9402,  26.0887],
#         [ 60.9402,  85.6474,  33.3293],
#         [ 26.0887,  33.3293, 104.0980]])

Evaluation

Metrics

Sparse Binary Classification

Dataset: quora_duplicates_dev
Evaluated with SparseBinaryClassificationEvaluator

Metric	Value
cosine_accuracy	0.759
cosine_accuracy_threshold	0.8013
cosine_f1	0.6742
cosine_f1_threshold	0.5425
cosine_precision	0.5282
cosine_recall	0.9317
cosine_ap	0.6876
cosine_mcc	0.506
dot_accuracy	0.754
dot_accuracy_threshold	47.2765
dot_f1	0.676
dot_f1_threshold	40.9553
dot_precision	0.5399
dot_recall	0.9037
dot_ap	0.6071
dot_mcc	0.5042
euclidean_accuracy	0.677
euclidean_accuracy_threshold	-14.2952
euclidean_f1	0.486
euclidean_f1_threshold	-0.5385
euclidean_precision	0.3213
euclidean_recall	0.9969
euclidean_ap	0.2043
euclidean_mcc	-0.0459
manhattan_accuracy	0.677
manhattan_accuracy_threshold	-163.6865
manhattan_f1	0.486
manhattan_f1_threshold	-2.7509
manhattan_precision	0.3213
manhattan_recall	0.9969
manhattan_ap	0.2056
manhattan_mcc	-0.0459
max_accuracy	0.759
max_accuracy_threshold	47.2765
max_f1	0.676
max_f1_threshold	40.9553
max_precision	0.5399
max_recall	0.9969
max_ap	0.6876
max_mcc	0.506
active_dims	83.3634
sparsity_ratio	0.9973

Sparse Information Retrieval

Datasets: NanoMSMARCO, NanoNQ, NanoNFCorpus, NanoQuoraRetrieval, NanoClimateFEVER, NanoDBPedia, NanoFEVER, NanoFiQA2018, NanoHotpotQA, NanoMSMARCO, NanoNFCorpus, NanoNQ, NanoQuoraRetrieval, NanoSCIDOCS, NanoArguAna, NanoSciFact and NanoTouche2020
Evaluated with SparseInformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNQ	NanoNFCorpus	NanoQuoraRetrieval	NanoClimateFEVER	NanoDBPedia	NanoFEVER	NanoFiQA2018	NanoHotpotQA	NanoSCIDOCS	NanoArguAna	NanoSciFact	NanoTouche2020
dot_accuracy@1	0.24	0.18	0.3	0.9	0.14	0.56	0.64	0.2	0.8	0.34	0.08	0.44	0.4082
dot_accuracy@3	0.44	0.46	0.42	0.96	0.32	0.78	0.72	0.28	0.9	0.56	0.32	0.58	0.7551
dot_accuracy@5	0.6	0.5	0.48	1.0	0.42	0.82	0.82	0.4	0.92	0.66	0.38	0.7	0.8776
dot_accuracy@10	0.74	0.64	0.52	1.0	0.52	0.88	0.88	0.46	0.94	0.78	0.44	0.78	0.9592
dot_precision@1	0.24	0.18	0.3	0.9	0.14	0.56	0.64	0.2	0.8	0.34	0.08	0.44	0.4082
dot_precision@3	0.1467	0.1533	0.2467	0.3867	0.1133	0.5133	0.2533	0.1267	0.3933	0.26	0.1067	0.2	0.4354
dot_precision@5	0.12	0.1	0.216	0.256	0.092	0.488	0.176	0.104	0.264	0.2	0.076	0.148	0.3837
dot_precision@10	0.074	0.066	0.174	0.136	0.064	0.436	0.094	0.07	0.142	0.142	0.044	0.086	0.3327
dot_recall@1	0.24	0.17	0.0201	0.804	0.0717	0.0423	0.6067	0.0947	0.4	0.0717	0.08	0.415	0.0271
dot_recall@3	0.44	0.43	0.0352	0.9087	0.1483	0.118	0.7033	0.1508	0.59	0.1607	0.32	0.55	0.0847
dot_recall@5	0.6	0.46	0.0744	0.97	0.19	0.1751	0.8033	0.2536	0.66	0.2057	0.38	0.665	0.1209
dot_recall@10	0.74	0.61	0.0892	0.99	0.25	0.2739	0.8633	0.3212	0.71	0.2917	0.44	0.76	0.2134
dot_ndcg@10	0.4666	0.3928	0.2175	0.9434	0.1928	0.5024	0.7369	0.2333	0.6849	0.285	0.2651	0.5848	0.3661
dot_mrr@10	0.3822	0.3355	0.3754	0.94	0.2527	0.6802	0.7064	0.2714	0.8542	0.4741	0.2085	0.54	0.5941
dot_map@100	0.3914	0.3266	0.0833	0.921	0.1415	0.3822	0.6976	0.1839	0.6061	0.2007	0.2135	0.5247	0.2483
query_active_dims	94.9	85.72	101.92	87.4	102.34	79.8	104.22	89.74	111.24	113.78	202.02	102.48	97.3061
query_sparsity_ratio	0.9969	0.9972	0.9967	0.9971	0.9966	0.9974	0.9966	0.9971	0.9964	0.9963	0.9934	0.9966	0.9968
corpus_active_dims	115.977	156.1067	217.0911	90.3262	217.8072	146.6807	228.7436	131.3409	166.1906	226.2181	176.6116	216.6451	147.0164
corpus_sparsity_ratio	0.9962	0.9949	0.9929	0.997	0.9929	0.9952	0.9925	0.9957	0.9946	0.9926	0.9942	0.9929	0.9952

Sparse Nano BEIR

Dataset: NanoBEIR_mean
Evaluated with SparseNanoBEIREvaluator with these parameters:
json
1{ 2 "dataset_names": [ 3 "msmarco", 4 "nq", 5 "nfcorpus", 6 "quoraretrieval" 7 ] 8}

Metric	Value
dot_accuracy@1	0.4
dot_accuracy@3	0.565
dot_accuracy@5	0.625
dot_accuracy@10	0.71
dot_precision@1	0.4
dot_precision@3	0.23
dot_precision@5	0.166
dot_precision@10	0.1075
dot_recall@1	0.306
dot_recall@3	0.4471
dot_recall@5	0.504
dot_recall@10	0.5844
dot_ndcg@10	0.4927
dot_mrr@10	0.5017
dot_map@100	0.4251
query_active_dims	83.545
query_sparsity_ratio	0.9973
corpus_active_dims	123.2832
corpus_sparsity_ratio	0.996

Sparse Nano BEIR

Dataset: NanoBEIR_mean
Evaluated with SparseNanoBEIREvaluator with these parameters:
json
1{ 2 "dataset_names": [ 3 "climatefever", 4 "dbpedia", 5 "fever", 6 "fiqa2018", 7 "hotpotqa", 8 "msmarco", 9 "nfcorpus", 10 "nq", 11 "quoraretrieval", 12 "scidocs", 13 "arguana", 14 "scifact", 15 "touche2020" 16 ] 17}

Metric	Value
dot_accuracy@1	0.4022
dot_accuracy@3	0.5765
dot_accuracy@5	0.6598
dot_accuracy@10	0.7338
dot_precision@1	0.4022
dot_precision@3	0.2566
dot_precision@5	0.2018
dot_precision@10	0.1431
dot_recall@1	0.2341
dot_recall@3	0.3569
dot_recall@5	0.4275
dot_recall@10	0.5041
dot_ndcg@10	0.4517
dot_mrr@10	0.5088
dot_map@100	0.3785
query_active_dims	105.6179
query_sparsity_ratio	0.9965
corpus_active_dims	163.7364
corpus_sparsity_ratio	0.9946

Training Details

Training Dataset

quora-duplicates

Dataset: quora-duplicates at 451a485
Size: 99,000 training samples
Columns: anchor, positive, and negative
Approximate statistics based on the first 1000 samples:
anchor positive negative
type string string string
details
min: 6 tokens
mean: 14.1 tokens
max: 39 tokens
min: 6 tokens
mean: 13.83 tokens
max: 41 tokens
min: 6 tokens
mean: 15.21 tokens
max: 75 tokens

Samples:

anchor	positive	negative
`What are the best GMAT coaching institutes in Delhi NCR?`	`Which are the best GMAT coaching institutes in Delhi/NCR?`	`What are the best GMAT coaching institutes in Delhi-Noida Area?`
`Is a third world war coming?`	`Is World War 3 more imminent than expected?`	`Since the UN is unable to control terrorism and groups like ISIS, al-Qaeda and countries that promote terrorism (even though it consumed those countries), can we assume that the world is heading towards World War III?`
`Should I build iOS or Android apps first?`	`Should people choose Android or iOS first to build their App?`	`How much more effort is it to build your app on both iOS and Android?`

Loss: SpladeLoss with these parameters:
json
1{ 2 "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score')", 3 "lambda_corpus": 3e-05, 4 "lambda_query": 5e-05 5}

Evaluation Dataset

quora-duplicates

Dataset: quora-duplicates at 451a485
Size: 1,000 evaluation samples
Columns: anchor, positive, and negative
Approximate statistics based on the first 1000 samples:
anchor positive negative
type string string string
details
min: 6 tokens
mean: 14.05 tokens
max: 40 tokens
min: 6 tokens
mean: 14.14 tokens
max: 44 tokens
min: 6 tokens
mean: 14.56 tokens
max: 60 tokens

Samples:

anchor	positive	negative
`What happens if we use petrol in diesel vehicles?`	`Why can't we use petrol in diesel?`	`Why are diesel engines noisier than petrol engines?`
`Why is Saltwater taffy candy imported in Switzerland?`	`Why is Saltwater taffy candy imported in Laos?`	`Is salt a consumer product?`
`Which is your favourite film in 2016?`	`What movie is the best movie of 2016?`	`What will the best movie of 2017 be?`

Loss: SpladeLoss with these parameters:
json
1{ 2 "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score')", 3 "lambda_corpus": 3e-05, 4 "lambda_query": 5e-05 5}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 12
per_device_eval_batch_size: 12
learning_rate: 2e-05
num_train_epochs: 1
bf16: True
load_best_model_at_end: True
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 12
per_device_eval_batch_size: 12
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: 0.0
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: 42
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
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: no_duplicates
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	quora_duplicates_dev_max_ap	NanoMSMARCO_dot_ndcg@10	NanoNQ_dot_ndcg@10	NanoNFCorpus_dot_ndcg@10	NanoQuoraRetrieval_dot_ndcg@10	NanoBEIR_mean_dot_ndcg@10	NanoClimateFEVER_dot_ndcg@10	NanoDBPedia_dot_ndcg@10	NanoFEVER_dot_ndcg@10	NanoFiQA2018_dot_ndcg@10	NanoHotpotQA_dot_ndcg@10	NanoSCIDOCS_dot_ndcg@10	NanoArguAna_dot_ndcg@10	NanoSciFact_dot_ndcg@10	NanoTouche2020_dot_ndcg@10
0.0242	200	6.2275	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0485	400	0.4129	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0727	600	0.3238	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0970	800	0.2795	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1212	1000	0.255	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1455	1200	0.2367	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1697	1400	0.25	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1939	1600	0.2742	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2	1650	-	0.1914	0.6442	0.3107	0.2820	0.1991	0.8711	0.4157	-	-	-	-	-	-	-	-	-
0.2182	1800	0.2102	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2424	2000	0.1797	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2667	2200	0.2021	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2909	2400	0.1734	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3152	2600	0.1849	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3394	2800	0.1871	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3636	3000	0.1685	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3879	3200	0.1512	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4	3300	-	0.1139	0.6637	0.4200	0.3431	0.1864	0.9222	0.4679	-	-	-	-	-	-	-	-	-
0.4121	3400	0.1165	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4364	3600	0.1518	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4606	3800	0.1328	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4848	4000	0.1098	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5091	4200	0.1389	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5333	4400	0.1224	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5576	4600	0.09	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5818	4800	0.1162	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6	4950	-	0.0784	0.6666	0.4404	0.3688	0.2239	0.9478	0.4952	-	-	-	-	-	-	-	-	-
0.6061	5000	0.1054	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6303	5200	0.0949	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6545	5400	0.1315	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6788	5600	0.1246	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7030	5800	0.1047	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7273	6000	0.0861	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7515	6200	0.103	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7758	6400	0.1062	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8	6600	0.1275	0.0783	0.6856	0.4666	0.3928	0.2175	0.9434	0.5051	-	-	-	-	-	-	-	-	-
0.8242	6800	0.1131	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8485	7000	0.0651	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8727	7200	0.0657	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8970	7400	0.1065	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9212	7600	0.0691	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9455	7800	0.1136	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9697	8000	0.0834	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9939	8200	0.0867	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
1.0	8250	-	0.0720	0.6876	0.4688	0.3711	0.1901	0.9408	0.4927	-	-	-	-	-	-	-	-	-
-1	-1	-	-	-	0.4666	0.3928	0.2175	0.9434	0.4517	0.1928	0.5024	0.7369	0.2333	0.6849	0.2850	0.2651	0.5848	0.3661

The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

Energy Consumed: 0.075 kWh
Carbon Emitted: 0.029 kg of CO2
Hours Used: 0.306 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: 4.2.0.dev0
Transformers: 4.52.4
PyTorch: 2.6.0+cu124
Accelerate: 1.5.1
Datasets: 2.21.0
Tokenizers: 0.21.1

Citation

BibTeX

Sentence Transformers

bibtex

@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",
}

SpladeLoss

bibtex

@misc{formal2022distillationhardnegativesampling,
      title={From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective},
      author={Thibault Formal and Carlos Lassance and Benjamin Piwowarski and Stéphane Clinchant},
      year={2022},
      eprint={2205.04733},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2205.04733},
}

SparseMultipleNegativesRankingLoss

bibtex

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

FlopsLoss

bibtex

@article{paria2020minimizing,
    title={Minimizing flops to learn efficient sparse representations},
    author={Paria, Biswajit and Yeh, Chih-Kuan and Yen, Ian EH and Xu, Ning and Ravikumar, Pradeep and P{'o}czos, Barnab{'a}s},
    journal={arXiv preprint arXiv:2004.05665},
    year={2020}
    }

	anchor	positive	negative
type	string	string	string
details	min: 6 tokens mean: 14.1 tokens max: 39 tokens	min: 6 tokens mean: 13.83 tokens max: 41 tokens	min: 6 tokens mean: 15.21 tokens max: 75 tokens