SentenceTransformer based on allenai/scibert_scivocab_uncased

This is a sentence-transformers model finetuned from allenai/scibert_scivocab_uncased. It maps sentences & paragraphs to a 768-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: allenai/scibert_scivocab_uncased
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 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': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

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("ML5562/fine-tuned-scibert_scivocab_uncased-all-json-M1_f16")
# Run inference
sentences = [
    'Select which statements are true regarding SCFGs.A penalty will be applied for any incorrect answers.',
    'The true statements regarding Stochastic Context-Free Grammars (SCFGs) are:\n\nA: The sum over all the probabilities of the rules of a SCFG that share the same left-hand side should sum up to 1, which is a fundamental property of probabilistic grammars. \nB: The probabilities of lexical rules of a SCFG correspond to emission probabilities of Hidden Markov Models (HMMs) for Part-of-Speech (PoS) tagging, indicating a similarity in how both types of models handle probabilities associated with observed events. \n\nThe other statements either misrepresent SCFG properties or are incorrect.',
    'The true statements regarding SCFGs are A and B. \n\nReason: A is true because the probabilities of rules with the same left-hand side must sum to 1 to maintain valid probability distributions. B is also true because lexical rules in SCFGs represent the probabilities of generating terminal symbols, analogous to emission probabilities in Hidden Markov Models (HMMs) used for Part-of-Speech (PoS) tagging. The other statements are either incorrect or not universally applicable to all SCFGs.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Triplet

Dataset: val-eval
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.6229

Training Details

Training Dataset

Unnamed Dataset

Size: 19,392 training samples
Columns: sentence_0, sentence_1, and sentence_2

Approximate statistics based on the first 1000 samples:

	sentence_0	sentence_1	sentence_2
type	string	string	string
details	min: 5 tokens mean: 91.5 tokens max: 512 tokens	min: 3 tokens mean: 348.13 tokens max: 512 tokens	min: 3 tokens mean: 318.16 tokens max: 512 tokens

Samples:

sentence_0	sentence_1	sentence_2
`In class, we saw Karger's beautiful randomized algorithm for finding a min-cut in an undirected graph $G=(V,E)$ with $n =`	V	$ vertices. Each iteration of Karger's algorithm can be implemented in time $O(n^2)$, and if repeated $\Theta(n^2 \log n)$ times, Karger's algorithm returns a min-cut with probability at least $1-1/n$. However, this leads to the often prohibitively large running time of $O(n^4 \log n)$. Karger and Stein made a crucial observation that allowed them to obtain a much faster algorithm for min-cut: the Karger-Stein algorithm runs in time $O(n^2 \log^3 n)$ and finds a min-cut with probability at least $1-1/n$. Explain in a couple of sentences the main idea that allowed Karger and Stein to modify Karger's algorithm into the much faster Karger-Stein algorithm. In other words, what are the main differences between the two algorithms?
`If we need to create a channel that protects confidentiality and we have at our disposal a channel that protects integrity and authenticity, we need to use`	Answer: 3 To protect confidentiality, we need to ensure that the information being sent over the channel cannot be read by unauthorized parties. The options provided suggest different methods that can be employed for security: 1. Symmetric key encryption: This method encrypts data using the same key for both encryption and decryption. While it does provide confidentiality, the question specifies that we already have a channel that protects integrity and authenticity, which might imply that we are looking for a method that can be integrated with that existing channel. 2. Message authentication codes (MACs): These are used to ensure the integrity and authenticity of a message but do not provide confidentiality. Therefore, this option does not address the need for confidentiality. 3. Public key encryption: This method uses a pair of keys (public and private) for encryption and decryption. It allows for secure transmission of data, ensuring confidentiality, especially w...	`Answer: 3 To protect confidentiality, public key encryption is necessary as it allows for secure data transmission while ensuring that only authorized parties can decrypt the message. This method complements the existing channel that protects integrity and authenticity, thereby addressing the confidentiality requirement effectively.`
`For a $n$-bit block cipher with $k$-bit key, given a plaintext-ciphertext pair, a key exhaustive search has an average number of trials of \dots`	To determine the average number of trials required for a key exhaustive search on a block cipher, we need to consider the following: 1. Key Space: A block cipher with a $k$-bit key has a total of $2^k$ possible keys. 2. Exhaustive Search: In an exhaustive search, one tries each possible key until the correct one is found. On average, the correct key will be found after trying half of the total keys. Therefore, in a key exhaustive search: - The average number of trials is given by: [ \text{Average Trials} = \frac{2^k}{2} = 2^{k-1} ] However, in terms of options provided, we're looking for what corresponds to the average trials. 3. Since the options include $2^k$ and $\frac{2^k + 1}{2}$, we consider that on average, we would try about half of the keyspace, which can be represented as $\frac{2^k + 1}{2}$ for approximation in the context of average calculations. Thus, the correct answer is: Answer: $\frac{2^k + 1}{2}$	To determine the average number of trials for a key exhaustive search in the context of a block cipher, we need to analyze the options given and the definitions involved. 1. In an $n$-bit block cipher, the number of possible keys is $2^k$ where $k$ is the bit length of the key. 2. An exhaustive search means testing every possible key until the correct one is found. 3. Since there are $2^k$ possible keys, in the worst case, we would need to try all $2^k$ keys. However, on average, if you were to randomly guess a key, you would expect to find the correct key after trying about half of all possible keys. Thus, the average number of trials for an exhaustive search would be: [ \text{Average trials} = \frac{2^k}{2} = \frac{2^k + 1}{2} ] This matches one of the options provided. Final Answer: 3

Loss: TripletLoss with these parameters:

{
    "distance_metric": "TripletDistanceMetric.EUCLIDEAN",
    "triplet_margin": 5
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 4
per_device_eval_batch_size: 4
num_train_epochs: 20
fp16: True
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: 4
per_device_eval_batch_size: 4
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: 20
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: True
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}
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: round_robin

Training Logs

Epoch	Step	Training Loss	val-eval_cosine_accuracy
0.1031	500	4.7355	0.5606
0.2063	1000	4.5245	0.5852
0.3094	1500	4.4665	0.5988
0.4125	2000	4.6664	0.5545
0.5157	2500	4.7732	0.5961
0.6188	3000	4.3502	0.5827
0.7219	3500	4.5098	0.5821
0.8251	4000	4.3916	0.5969
0.9282	4500	4.5026	0.5965
1.0	4848	-	0.6106
1.0314	5000	4.3997	0.6118
1.1345	5500	4.131	0.5992
1.2376	6000	4.005	0.6038
1.3408	6500	4.0346	0.5990
1.4439	7000	4.1737	0.5959
1.5470	7500	4.256	0.6048
1.6502	8000	4.035	0.6122
1.7533	8500	4.0693	0.6083
1.8564	9000	4.2146	0.5978
1.9596	9500	4.3037	0.6229

Framework Versions

Python: 3.12.8
Sentence Transformers: 3.4.1
Transformers: 4.48.2
PyTorch: 2.5.1+cu124
Accelerate: 1.3.0
Datasets: 3.2.0
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",
}

TripletLoss

@misc{hermans2017defense,
    title={In Defense of the Triplet Loss for Person Re-Identification},
    author={Alexander Hermans and Lucas Beyer and Bastian Leibe},
    year={2017},
    eprint={1703.07737},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

ML5562
/

fine-tuned-scibert_scivocab_uncased-all-json-M1_f16