metadata

tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:400
  - loss:MatryoshkaLoss
  - loss:MultipleNegativesRankingLoss
base_model: Snowflake/snowflake-arctic-embed-l
widget:
  - source_sentence: >-
      What is the title of the dataset introduced by Jin, B. Dhingra, Z. Liu, W.
      Cohen, and X. Lu in their 2019 publication?
    sentences:
      - TechQA [3]
      - "Q.\_Jin, B.\_Dhingra, Z.\_Liu, W.\_Cohen, and X.\_Lu.\n\n\nPubMedQA: A dataset for biomedical research question answering.\n\n\nIn K.\_Inui, J.\_Jiang, V.\_Ng, and X.\_Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567–2577, Hong Kong, China, Nov. 2019. Association for Computational Linguistics.\n\n\ndoi: 10.18653/v1/D19-1259."
      - "Our contributions address the need for standardized benchmarks and methodologies, enabling more precise and actionable insights into the strengths and weaknesses of different RAG systems. This, in turn, will facilitate iterative improvement of RAG models, driving forward the capabilities of retrieval-augmented generation in real-world applications.\n\n\n\nReferences\n\n\nAdlakha et\_al. [2023]\n\nV.\_Adlakha, P.\_BehnamGhader, X.\_H. Lu, N.\_Meade, and S.\_Reddy."
  - source_sentence: >-
      What does the 2024 paper by Es et al. propose regarding the evaluation of
      retrieval augmented generation?
    sentences:
      - "doi: 10.18653/v1/2023.findings-acl.60.\n\n\nURL https://aclanthology.org/2023.findings-acl.60.\n\n\n\n\nDinan et\_al. [2019]\n\nE.\_Dinan, S.\_Roller, K.\_Shuster, A.\_Fan, M.\_Auli, and J.\_Weston.\n\n\nWizard of wikipedia: Knowledge-powered conversational agents, 2019.\n\n\n\n\nEs et\_al. [2024]\n\nS.\_Es, J.\_James, L.\_Espinosa\_Anke, and S.\_Schockaert.\n\n\nRAGAs: Automated evaluation of retrieval augmented generation."
      - >-
        Source Domains


        RAGBench comprises five distinct domains: bio-medical research
        (PubmedQA, CovidQA), general knowledge (HotpotQA, MS Marco, HAGRID,
        ExperQA), legal contracts (CuAD), customer support (DelucionQA, EManual,
        TechQA), and finance (FinBench, TAT-QA). We select these specific
        domains based on availability of data, and applicability to real-world
        RAG applications across different industry verticals. For detailed
        descriptions of each component data source, refer to Appendix 7.2.
      - >-
        The overall_supported_explanation field is a string explaining why the
        response

        *as a whole* is or is not supported by the documents. In this field,
        provide a

        step-by-step breakdown of the claims made in the response and the
        support (or

        lack thereof) for those claims in the documents. Begin by assessing each
        claim

        separately, one by one; don’t make any remarks about the response as a
        whole

        until you have assessed all the claims in isolation.
  - source_sentence: What are some common sources of questions in research or surveys?
    sentences:
      - "Kwiatkowski et\_al. [2019]\n\nT.\_Kwiatkowski, J.\_Palomaki, O.\_Redfield, M.\_Collins, A.\_Parikh, C.\_Alberti, D.\_Epstein, I.\_Polosukhin, M.\_Kelcey, J.\_Devlin, K.\_Lee, K.\_N. Toutanova, L.\_Jones, M.-W. Chang, A.\_Dai, J.\_Uszkoreit, Q.\_Le, and S.\_Petrov.\n\n\nNatural questions: a benchmark for question answering research.\n\n\nTransactions of the Association of Computational Linguistics, 2019.\n\n\n\n\nLaurer et\_al. [2022]\n\nM.\_Laurer, W.\_van Atteveldt, A.\_Casas, and K.\_Welbers."
      - Question Sources
      - >-
        the overall RAG system performance, with the potential to provide
        granular, actionable insights to the RAG practitioner.
  - source_sentence: >-
      What evaluation metrics are reported for the response-level hallucination
      detection task?
    sentences:
      - >-
        4.3 Evaluation


        Our granular annotation schema allows for various evaluation setups. For
        example, we could evaluate either span-level or example/response-level
        predictions. For easy comparison with existing RAG evaluation approaches
        that are less granular, we report area under the receiver-operator curve
        (AUROC) on the response-level hallucination detection task, and root
        mean squared error (RMSE) for example-level context Relevance and
        Utilization predictions.
      - >-
        EManual is a question answer dataset comprising consumer electronic
        device manuals and realistic questions about them composed by human
        annotators. The subset made available at the time of writing amounts to
        659 unique questions about the Samsung Smart TV/remote and the
        accompanying user manual, segmented into 261 chunks. To form a RAG
        dataset, we embed the manual segments into a vector database with OpenAI
        embedding and retrieve up to 3 context documents per question from it.
        For each
      - >-
        Table 3: Benchmark evaluation on test splits. Reporting AUROC for
        predicting hallucinated responses (Hal), RMSE for predicting Context
        Relevance (Rel) and utilization (Util). ∗ indicates statistical
        significance at 95% confidence intervals, measured by bootstrap
        comparing the top and second-best results. RAGAS and Trulens do not
        evaluate Utilization.





        GPT-3.5

        RAGAS

        TruLens

        DeBERTA



        Dataset



        Hal↑↑\uparrow↑





        Rel↓↓\downarrow↓



        Util↓↓\downarrow↓




        Hal↑↑\uparrow↑





        Rel↓↓\downarrow↓
  - source_sentence: >-
      What is the main contribution of Kwiatkowski et al. [2019] in the field of
      question answering research?
    sentences:
      - >-
        The sentence_support_information field is a list of objects, one for
        each sentence

        in the response. Each object MUST have the following fields:

        - response_sentence_key: a string identifying the sentence in the
        response.

        This key is the same as the one used in the response above.

        - explanation: a string explaining why the sentence is or is not
        supported by the

        documents.

        - supporting_sentence_keys: keys (e.g. ’0a’) of sentences from the
        documents that
      - "Kwiatkowski et\_al. [2019]\n\nT.\_Kwiatkowski, J.\_Palomaki, O.\_Redfield, M.\_Collins, A.\_Parikh, C.\_Alberti, D.\_Epstein, I.\_Polosukhin, M.\_Kelcey, J.\_Devlin, K.\_Lee, K.\_N. Toutanova, L.\_Jones, M.-W. Chang, A.\_Dai, J.\_Uszkoreit, Q.\_Le, and S.\_Petrov.\n\n\nNatural questions: a benchmark for question answering research.\n\n\nTransactions of the Association of Computational Linguistics, 2019.\n\n\n\n\nLaurer et\_al. [2022]\n\nM.\_Laurer, W.\_van Atteveldt, A.\_Casas, and K.\_Welbers."
      - >-
        with consistent annotations. To best represent real-world RAG scenarios,
        we vary a number parameters to construct the benchmark: the source
        domain, number of context documents, context token length, and the
        response generator model Figure 1 illustrates where these variable
        parameters fall in the RAG pipeline.
pipeline_tag: sentence-similarity
library_name: sentence-transformers
metrics:
  - cosine_accuracy@1
  - cosine_accuracy@3
  - cosine_accuracy@5
  - cosine_accuracy@10
  - cosine_precision@1
  - cosine_precision@3
  - cosine_precision@5
  - cosine_precision@10
  - cosine_recall@1
  - cosine_recall@3
  - cosine_recall@5
  - cosine_recall@10
  - cosine_ndcg@10
  - cosine_mrr@10
  - cosine_map@100
model-index:
  - name: SentenceTransformer based on Snowflake/snowflake-arctic-embed-l
    results:
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: Unknown
          type: unknown
        metrics:
          - type: cosine_accuracy@1
            value: 0.8571428571428571
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.9642857142857143
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 1
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 1
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.8571428571428571
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.32142857142857145
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.20000000000000004
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.10000000000000002
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.8571428571428571
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.9642857142857143
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 1
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 1
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.9385586452838898
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.9178571428571428
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.9178571428571428
            name: Cosine Map@100

SentenceTransformer based on Snowflake/snowflake-arctic-embed-l

This is a sentence-transformers model finetuned from Snowflake/snowflake-arctic-embed-l. It maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: Snowflake/snowflake-arctic-embed-l
Maximum Sequence Length: 512 tokens
Output Dimensionality: 1024 dimensions
Similarity Function: Cosine Similarity

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

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 SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("chelleboyer/llm-evals-2-79b954ef-4798-4994-be72-a88d46b8ecca")
# Run inference
sentences = [
    'What is the main contribution of Kwiatkowski et al. [2019] in the field of question answering research?',
    'Kwiatkowski et\xa0al. [2019]\n\nT.\xa0Kwiatkowski, J.\xa0Palomaki, O.\xa0Redfield, M.\xa0Collins, A.\xa0Parikh, C.\xa0Alberti, D.\xa0Epstein, I.\xa0Polosukhin, M.\xa0Kelcey, J.\xa0Devlin, K.\xa0Lee, K.\xa0N. Toutanova, L.\xa0Jones, M.-W. Chang, A.\xa0Dai, J.\xa0Uszkoreit, Q.\xa0Le, and S.\xa0Petrov.\n\n\nNatural questions: a benchmark for question answering research.\n\n\nTransactions of the Association of Computational Linguistics, 2019.\n\n\n\n\nLaurer et\xa0al. [2022]\n\nM.\xa0Laurer, W.\xa0van Atteveldt, A.\xa0Casas, and K.\xa0Welbers.',
    'The sentence_support_information field is a list of objects, one for each sentence\nin the response. Each object MUST have the following fields:\n- response_sentence_key: a string identifying the sentence in the response.\nThis key is the same as the one used in the response above.\n- explanation: a string explaining why the sentence is or is not supported by the\ndocuments.\n- supporting_sentence_keys: keys (e.g. ’0a’) of sentences from the documents that',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Information Retrieval

Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.8571
cosine_accuracy@3	0.9643
cosine_accuracy@5	1.0
cosine_accuracy@10	1.0
cosine_precision@1	0.8571
cosine_precision@3	0.3214
cosine_precision@5	0.2
cosine_precision@10	0.1
cosine_recall@1	0.8571
cosine_recall@3	0.9643
cosine_recall@5	1.0
cosine_recall@10	1.0
cosine_ndcg@10	0.9386
cosine_mrr@10	0.9179
cosine_map@100	0.9179

Training Details

Training Dataset

Unnamed Dataset

Size: 400 training samples
Columns: sentence_0 and sentence_1
Approximate statistics based on the first 400 samples:
sentence_0 sentence_1
type string string
details
min: 3 tokens
mean: 21.42 tokens
max: 53 tokens

min: 3 tokens
mean: 93.8 tokens
max: 200 tokens

	sentence_0	sentence_1
type	string	string
details	min: 3 tokens mean: 21.42 tokens max: 53 tokens	min: 3 tokens mean: 93.8 tokens max: 200 tokens

Samples:

sentence_0	sentence_1
`What are the key components and criteria used in the TRACe Evaluation Framework within RAGBench?`	`RAGBench: Explainable Benchmark for Retrieval-Augmented Generation Systems 1 Introduction 2 Related Work RAG evaluation Finetuned RAG evaluation models 3 RAGBench Construction 3.1 Component Datasets Source Domains Context Token Length Task Types Question Sources Response Generation Data Splits 3.2 TRACe Evaluation Framework Definitions Context Relevance Context Utilization Completeness Adherence 3.3 RAGBench Statistics 3.4 LLM annotator`
`How does RAGBench utilize component datasets to construct a benchmark for Retrieval-Augmented Generation systems?`	`RAGBench: Explainable Benchmark for Retrieval-Augmented Generation Systems 1 Introduction 2 Related Work RAG evaluation Finetuned RAG evaluation models 3 RAGBench Construction 3.1 Component Datasets Source Domains Context Token Length Task Types Question Sources Response Generation Data Splits 3.2 TRACe Evaluation Framework Definitions Context Relevance Context Utilization Completeness Adherence 3.3 RAGBench Statistics 3.4 LLM annotator`
`What are the key components and findings discussed in the RAGBench Statistics and Case Study sections?`	`3.3 RAGBench Statistics 3.4 LLM annotator Alignment with Human Judgements 3.5 RAG Case Study 4 Experiments 4.1 LLM Judge 4.2 Fine-tuned Judge 4.3 Evaluation 5 Results Estimating Context Relevance is Difficult 6 Conclusion 7 Appendix 7.1 RAGBench Code and Data 7.2 RAGBench Dataset Details PubMedQA [14] CovidQA-RAG HotpotQA [42] MS Marco [28] CUAD [12] DelucionQA [33] EManual [27] TechQA [3] FinQA [6] TAT-QA [47] HAGRID [15] ExpertQA [25]`

Loss: MatryoshkaLoss with these parameters:

{
    "loss": "MultipleNegativesRankingLoss",
    "matryoshka_dims": [
        768,
        512,
        256,
        128,
        64
    ],
    "matryoshka_weights": [
        1,
        1,
        1,
        1,
        1
    ],
    "n_dims_per_step": -1
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 5
per_device_eval_batch_size: 5
num_train_epochs: 10
multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 5
per_device_eval_batch_size: 5
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: 5e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1
num_train_epochs: 10
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: False
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: False
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: round_robin

Training Logs

Epoch	Step	Training Loss	cosine_ndcg@10
0.625	50	-	0.9517
1.0	80	-	0.9649
1.25	100	-	0.9649
1.875	150	-	0.9517
2.0	160	-	0.9517
2.5	200	-	0.9386
3.0	240	-	0.9386
3.125	250	-	0.9517
3.75	300	-	0.9386
4.0	320	-	0.9517
4.375	350	-	0.9517
5.0	400	-	0.9517
5.625	450	-	0.9517
6.0	480	-	0.9401
6.25	500	0.3877	0.9401
6.875	550	-	0.9386
7.0	560	-	0.9386
7.5	600	-	0.9401
8.0	640	-	0.9401
8.125	650	-	0.9401
8.75	700	-	0.9386
9.0	720	-	0.9386
9.375	750	-	0.9386
10.0	800	-	0.9386

Framework Versions

Python: 3.11.12
Sentence Transformers: 4.1.0
Transformers: 4.51.3
PyTorch: 2.6.0+cu124
Accelerate: 1.6.0
Datasets: 2.14.4
Tokenizers: 0.21.1

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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

@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}
}