mrm8488/distilroberta-base-ft-webinstruct

SentenceTransformer based on distilbert/distilroberta-base

This is a sentence-transformers model finetuned from distilbert/distilroberta-base on the mathstackexchange, socratic and stackexchange datasets. 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: distilbert/distilroberta-base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Datasets:
  • mathstackexchange
  • socratic
  • stackexchange

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: RobertaModel 
  (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("mrm8488/distilroberta-base-ft-webintruct-512")
# Run inference
sentences = [
    "What should I do if I'm not satisfied with the answers to a question for which I've offered a bounty?\n\nIn my case, I've put a bounty on a question, but the two responses I received don't address the issue effectively. I requested the original poster (OP) to provide an answer so I could reward them for the interesting question, but they haven't done so. \n\nAre there any acceptable actions in this scenario? For instance, can I post my own non-answer, award myself the bounty, and then start a new bounty on a different question? Or are there alternative suggestions?",
    "If all the provided answers do not adequately address your question, it's advisable to let the bounty expire. The system will handle the distribution of the bounty in such situations according to predefined rules.\n\nBounties carry a risk, as there is no guarantee that you will receive a satisfactory answer, even with the incentive. It's important to understand that you cannot reclaim your bounty once it's been offered. \n\nInstead of posting a non-answer, you might consider editing and clarifying your original question to attract better responses, or seeking assistance from the community through comments or chat. If needed, you can also start a new bounty on a different question, but ensure that it's clear and well-defined to increase the likelihood of receiving quality answers.",
    'The issue you\'re experiencing with your 40kHz crystal oscillator might be due to insufficient drive strength and an incorrect load capacitance. Here are two potential causes and solutions:\n\n1. High Series Resistance: The 150 kΩ series resistance in your circuit might be too high, which results in a low drive strength for the crystal. This can lead to a reduced overall loop gain and prevents the oscillator from properly starting. To resolve this, try using a lower resistance value as recommended in the crystal\'s datasheet.\n\n2. Incorrect Load Capacitance: Ensure that the 33 pF load capacitors you\'re using are compatible with your crystal. Some low-power "watch" crystals require only 5-10 pF load capacitors. Always refer to the crystal\'s datasheet to verify the appropriate load capacitance value.\n\nIn summary, carefully review the crystal\'s datasheet to determine the correct series resistance and load capacitance values, and make the necessary adjustments to your circuit. By doing so, you should be able to resolve the issue and get your oscillator functioning properly.',
]
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

Semantic Similarity

• Dataset: sts-dev
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.7294
spearman_cosine	0.7554
pearson_manhattan	0.7659
spearman_manhattan	0.7695
pearson_euclidean	0.767
spearman_euclidean	0.7704
pearson_dot	0.5262
spearman_dot	0.5063
pearson_max	0.767
spearman_max	0.7704

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.7064
spearman_cosine	0.7116
pearson_manhattan	0.7219
spearman_manhattan	0.7117
pearson_euclidean	0.7233
spearman_euclidean	0.7131
pearson_dot	0.4464
spearman_dot	0.4372
pearson_max	0.7233
spearman_max	0.7131

Training Details

Training Datasets

mathstackexchange

• Dataset: mathstackexchange
• Size: 1,484,629 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 10 tokens mean: 85.8 tokens max: 512 tokens	min: 12 tokens mean: 298.22 tokens max: 512 tokens

• Samples:

  anchor	positive
  Suppose $A$ is a normal subgroup of a group $B$, and the quotient group $B/A$ is cyclic with infinite order. How can we demonstrate, using the correspondence theorem, that for every positive integer $k$, $B$ has a normal subgroup of index $k$?	The correspondence theorem relates subgroups of the quotient group $B/A$ to subgroups of $B$ containing $A$. Since $B/A$ is isomorphic to the infinite cyclic group $\mathbb{Z}$, it has subgroups of every finite index. To find a normal subgroup of $B$ with index $k$, we can follow these steps: 1. Identify a subgroup $M/A$ of $B/A$ with index $k$. This is possible since $\mathbb{Z}$ has subgroups of index $k$ for any positive integer $k$ (for instance, the subgroup generated by $k$ times the generator of $\mathbb{Z}$). 2. By the correspondence theorem, there exists an intermediate subgroup $M$ such that $A \leq M \leq B$ with $M/A \cong M/A$. 3. Now, consider the index of $M$ in $B$: $[B:M] = [B/A : M/A] = k$, as desired. Thus, using the correspondence theorem and exploiting the properties of $\mathbb{Z}$, we can construct a normal subgroup of $B$ with index $k$ for any positive integer $k$.
  If $z$ is a complex number and $a$ is a real number, can we say $	az
  What is the remainder when $x^{2007}$ is divided by $x^2-x+1$?	To find the remainder when $x^{2007}$ is divided by $x^2-x+1$, we can use polynomial long division or synthetic division. Alternatively, we can utilize the properties of the polynomial $x^2-x+1$. Since $x^2-x+1$ is a factor of $x^3+1$, we have $x^3 \equiv -1 \pmod{x^2-x+1}$. Therefore, we can express $x^{2007}$ as $x^{3\cdot669}$, and thus: $$x^{2007} = (x^3)^{669} \equiv (-1)^{669} \pmod{x^2-x+1}$$ Now, since $669$ is odd, $(-1)^{669} = -1$. Hence, the remainder is $-1$. Alternatively, we can perform polynomial long division to obtain: $$\frac{x^{2007}}{x^2-x+1} = a(x) - \frac{x+1}{x^2-x+1}$$ where $a(x)$ is the quotient polynomial. Since we only care about the remainder, we have: $$x^{2007} \equiv -1 \pmod{x^2-x+1}$$

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

socratic

• Dataset: socratic
• Size: 533,383 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 8 tokens mean: 29.99 tokens max: 156 tokens	min: 3 tokens mean: 210.0 tokens max: 512 tokens

• Samples:

  anchor	positive
  What is an activated complex?	The activated complex is formed when the reactants collide with each other and begin to rearrange their atoms and bonds to form the products. This process requires energy, which is why the activated complex has a higher energy than the reactants. The energy required to reach the activated complex is called the activation energy. Once the activated complex is formed, it can either decompose back into the reactants or proceed to form the products. The probability of the activated complex decomposing back into the reactants is determined by the activation energy. If the activation energy is high, then the activated complex is more likely to decompose back into the reactants. If the activation energy is low, then the activated complex is more likely to proceed to form the products. The activated complex is a key concept in understanding chemical reactions. It helps to explain why some reactions occur quickly and others occur slowly. It also helps to explain why some reactions require a catalyst to occur. #### An activated complex is a high-energy, unstable intermediate state that forms during a chemical reaction. It is the transition state between the reactants and the products. The activated complex has a higher energy than both the reactants and the products, and it represents the maximum energy that must be overcome for the reaction to occur.
  Why does gravity cause planets to be round?	The gravitational force of a planet pulls matter towards its center. This force is strongest at the center of the planet and weakest at the surface. As a result, matter is pulled towards the center of the planet, causing it to take on a spherical shape. A sphere is the shape with the lowest surface area for a given volume. This means that a planet with a spherical shape has the least amount of potential energy. Potential energy is the energy that an object has due to its position or condition. In the case of a planet, its potential energy is due to its gravitational force. The lower the potential energy of a planet, the more stable it is. This is because a planet with a lower potential energy is less likely to change its shape. As a result, planets tend to be spherical in shape. #### Gravity causes planets to be round because a sphere is the shape with the lowest surface area for a given volume. This means that a planet with a spherical shape has the least amount of potential energy, which is the energy that an object has due to its position or condition. Gravity pulls matter towards the center of a planet, and this force is strongest at the center. As a result, matter is pulled towards the center of the planet, causing it to take on a spherical shape.
  How many carbon atoms are present in a 5.85-gram sample of carbon tetrabromide (CBr4)?	There are approximately (1 \times 10^{22}) carbon atoms in the given sample. Explanation: To determine the number of carbon atoms, we first need to calculate the molar amount of CBr4. The molar mass of CBr4 is 331.63 g/mol. So, we have: [ \frac{5.85\ g}{331.63\ g/mol} = 0.0176\ mol ] Since one molecule of CBr4 contains one carbon atom and four bromine atoms, there are: [ 1 \times 0.0176\ mol = 0.0176\ mol\ of\ carbon\ atoms ] Now, multiplying the molar quantity by Avogadro's number (6.022 × 10^23 mol^(-1)) gives us the number of individual carbon atoms: [ 0.0176\ mol \times 6.022 \times 10^{23}\ mol^{-1} = 1.06 \times 10^{22}\ carbon\ atoms ] Therefore, there are approximately (1 \times 10^{22}) carbon atoms in a 5.85-gram sample of CBr4.

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

stackexchange

• Dataset: stackexchange
• Size: 317,208 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 12 tokens mean: 63.36 tokens max: 512 tokens	min: 50 tokens mean: 263.31 tokens max: 512 tokens

• Samples:

  anchor	positive
  Should I use a tip activator to recoat the worn protective coating on my iron tip, or is it better to replace the tip entirely? My 48W ZD99 Solder Station's tip is showing signs of peeling due to moisture exposure and inadequate care. Can the tip activator effectively restore the tip, or should I opt for a new one?	To address the issue, first clean the iron tip with a wire brush to remove any debris. Then, apply flux and tin the tip to protect it and maintain its performance. Tip activators are available as a means to recoat tips, but their effectiveness may vary. While they can be a viable solution, it's essential to ensure proper tip care to prevent future wear. If the tip's condition significantly deteriorates despite these efforts, consider replacing it with a new one.
  What are the fundamental limits, if any, for the speed of sound in different materials, and how do these limits relate to the speed of light?	The speed of sound is limited by the properties of the material it travels through and the fundamental principles of physics. In a theoretical sense, the maximum speed of sound is constrained by the speed of light (approximately 299,792 km/s in vacuum), which is the maximum speed at which information can propagate. This limit is reached when the material has an incompressible equation of state, such as in the core of a neutron star, where the strong nuclear force creates immense pressure resistance. For an ideal gas, where particles do not interact, the equation of state is the softest possible with $P = \rho c^2/3$, where $P$ is pressure, $\rho$ is density, and $c$ is the speed of light. In this case, the maximum speed of sound would be $c/\sqrt{3}$. It's important to note that in practice, materials with extremely high sound speeds are unlikely to exist due to the conditions required for an incompressible equation of state. In reality, materials like solids and liquids generally have faster sound speeds than gases, but they are still far below the speed of light. When dealing with exotic materials, such as short-lived isotopes or neutron stars, the speed of sound may be even more challenging to determine due to the unique properties and states involved. However, the underlying principles remain the same: the speed of sound is determined by the material's properties, and it cannot exceed the speed of light in a vacuum.
  What could be causing a 1996 Honda Civic to stop running suddenly, and how can it be started?	A potential issue is a faulty ignition switch. When you attempt to start the car, the switch might be malfunctioning in such a way that it disrupts power to the engine ignition system, causing the dash lights to go out and preventing the car from starting. However, when you perform a push start (crash start), the car starts because the ignition switch remains in position 2, providing power to the engine. Another possibility is a problem with the battery or its connections. If the battery terminals have a poor connection, it might lead to high resistance, making it difficult for the car to start. Alternatively, if the battery is weak, it might not supply enough power to crank the engine effectively. In this case, the starter motor would sound sluggish as it tries to turn the engine. To resolve the issue, inspect the ignition switch, battery connections, and consider testing or replacing the battery if necessary.

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• num_train_epochs: 1
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates
• 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: 64
• per_device_eval_batch_size: 64
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_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: 1.0
• 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: 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: False
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: round_robin

Training Logs

Click to expand

Epoch	Step	Training Loss	sts-dev_spearman_cosine	sts-test_spearman_cosine
0.0067	100	3.6873	-	-
0.0134	200	0.984	-	-
0.0202	300	0.2259	-	-
0.0269	400	0.1696	-	-
0.0336	500	0.1468	-	-
0.0403	600	0.1235	-	-
0.0471	700	0.1125	-	-
0.0538	800	0.1032	-	-
0.0605	900	0.097	-	-
0.0672	1000	0.0992	0.8011	-
0.0740	1100	0.0937	-	-
0.0807	1200	0.0818	-	-
0.0874	1300	0.0909	-	-
0.0941	1400	0.0836	-	-
0.1009	1500	0.0705	-	-
0.1076	1600	0.081	-	-
0.1143	1700	0.0791	-	-
0.1210	1800	0.0677	-	-
0.1278	1900	0.0697	-	-
0.1345	2000	0.0661	0.7721	-
0.1412	2100	0.0727	-	-
0.1479	2200	0.0683	-	-
0.1547	2300	0.0597	-	-
0.1614	2400	0.06	-	-
0.1681	2500	0.0598	-	-
0.1748	2600	0.051	-	-
0.1816	2700	0.0629	-	-
0.1883	2800	0.0513	-	-
0.1950	2900	0.0517	-	-
0.2017	3000	0.048	0.7783	-
0.2085	3100	0.0418	-	-
0.2152	3200	0.0447	-	-
0.2219	3300	0.0458	-	-
0.2286	3400	0.0504	-	-
0.2354	3500	0.0463	-	-
0.2421	3600	0.0433	-	-
0.2488	3700	0.0447	-	-
0.2555	3800	0.0444	-	-
0.2623	3900	0.0432	-	-
0.2690	4000	0.0452	0.7910	-
0.2757	4100	0.0419	-	-
0.2824	4200	0.0373	-	-
0.2892	4300	0.0385	-	-
0.2959	4400	0.0381	-	-
0.3026	4500	0.0383	-	-
0.3093	4600	0.0367	-	-
0.3161	4700	0.0353	-	-
0.3228	4800	0.034	-	-
0.3295	4900	0.0333	-	-
0.3362	5000	0.0406	0.7862	-
0.3429	5100	0.0319	-	-
0.3497	5200	0.0332	-	-
0.3564	5300	0.0337	-	-
0.3631	5400	0.0347	-	-
0.3698	5500	0.0333	-	-
0.3766	5600	0.036	-	-
0.3833	5700	0.0319	-	-
0.3900	5800	0.0342	-	-
0.3967	5900	0.0296	-	-
0.4035	6000	0.0313	0.7675	-
0.4102	6100	0.0289	-	-
0.4169	6200	0.0292	-	-
0.4236	6300	0.0271	-	-
0.4304	6400	0.0295	-	-
0.4371	6500	0.0353	-	-
0.4438	6600	0.035	-	-
0.4505	6700	0.0324	-	-
0.4573	6800	0.0281	-	-
0.4640	6900	0.0265	-	-
0.4707	7000	0.031	0.7634	-
0.4774	7100	0.0302	-	-
0.4842	7200	0.0268	-	-
0.4909	7300	0.0275	-	-
0.4976	7400	0.0267	-	-
0.5043	7500	0.0249	-	-
0.5111	7600	0.0285	-	-
0.5178	7700	0.0311	-	-
0.5245	7800	0.0248	-	-
0.5312	7900	0.0278	-	-
0.5380	8000	0.0267	0.7658	-
0.5447	8100	0.0245	-	-
0.5514	8200	0.0261	-	-
0.5581	8300	0.0227	-	-
0.5649	8400	0.0261	-	-
0.5716	8500	0.0241	-	-
0.5783	8600	0.0261	-	-
0.5850	8700	0.0173	-	-
0.5918	8800	0.0226	-	-
0.5985	8900	0.0221	-	-
0.6052	9000	0.023	0.7558	-
0.6119	9100	0.0218	-	-
0.6187	9200	0.0245	-	-
0.6254	9300	0.0232	-	-
0.6321	9400	0.0208	-	-
0.6388	9500	0.0202	-	-
0.6456	9600	0.022	-	-
0.6523	9700	0.0212	-	-
0.6590	9800	0.0228	-	-
0.6657	9900	0.0214	-	-
0.6724	10000	0.0206	0.7686	-
0.6792	10100	0.0227	-	-
0.6859	10200	0.0225	-	-
0.6926	10300	0.018	-	-
0.6993	10400	0.0185	-	-
0.7061	10500	0.0204	-	-
0.7128	10600	0.0216	-	-
0.7195	10700	0.0212	-	-
0.7262	10800	0.0156	-	-
0.7330	10900	0.0232	-	-
0.7397	11000	0.0146	0.7610	-
0.7464	11100	0.0165	-	-
0.7531	11200	0.0187	-	-
0.7599	11300	0.0199	-	-
0.7666	11400	0.0215	-	-
0.7733	11500	0.0222	-	-
0.7800	11600	0.021	-	-
0.7868	11700	0.0163	-	-
0.7935	11800	0.0192	-	-
0.8002	11900	0.0206	-	-
0.8069	12000	0.017	0.7658	-
0.8137	12100	0.0152	-	-
0.8204	12200	0.0175	-	-
0.8271	12300	0.0211	-	-
0.8338	12400	0.0162	-	-
0.8406	12500	0.0178	-	-
0.8473	12600	0.0142	-	-
0.8540	12700	0.02	-	-
0.8607	12800	0.0166	-	-
0.8675	12900	0.0187	-	-
0.8742	13000	0.017	0.7603	-
0.8809	13100	0.0167	-	-
0.8876	13200	0.0211	-	-
0.8944	13300	0.0162	-	-
0.9011	13400	0.0161	-	-
0.9078	13500	0.0157	-	-
0.9145	13600	0.016	-	-
0.9213	13700	0.0139	-	-
0.9280	13800	0.0175	-	-
0.9347	13900	0.0172	-	-
0.9414	14000	0.0148	0.7554	-
0.9482	14100	0.0227	-	-
0.9549	14200	0.0174	-	-
0.9616	14300	0.0191	-	-
0.9683	14400	0.0151	-	-
0.9751	14500	0.0184	-	-
0.9818	14600	0.02	-	-
0.9885	14700	0.0163	-	-
0.9952	14800	0.0141	-	-
1.0	14871	-	-	0.7116

Framework Versions

• Python: 3.10.12
• Sentence Transformers: 3.0.1
• Transformers: 4.42.4
• PyTorch: 2.4.0+cu121
• Accelerate: 0.34.0
• Datasets: 2.21.0
• Tokenizers: 0.19.1

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",
}

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}
}