This is a sentence-transformers model trained on the parquet dataset. It maps sentences & paragraphs to a 512-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

Maximum Sequence Length: 512 tokens
Output Dimensionality: 512 dimensions
Similarity Function: Cosine Similarity
Training Dataset:
- parquet

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': 512, '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.

python

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("pankajrajdeo/Bioformer-8L-UMLS-Pubmed_PMC-Random_TCE-Epoch-1")
# Run inference
sentences = [
    'Development of an in vitro regeneration system from immature inflorescences and CRISPR/Cas9-mediated gene editing in sudangrass.',
    'BACKGROUND: Sudangrass (Sorghum sudanense) is a major biomass producer for livestock feed and biofuel in many countries. It has a wide range of adaptations for growing on marginal lands under biotic and abiotic stresses. The immature inflorescence is an explant with high embryogenic competence and is frequently used to regenerate different sorghum cultivars. Caffeic acid O-methyl transferase (COMT) is a key enzyme in the lignin biosynthesis pathway, which limits ruminant digestion of forage cell walls and is a crucial barrier in the conversion of plant biomass to bioethanol. Genome editing by CRISPR/Cas9-mediated mutagenesis without a transgenic footprint will accelerate the improvement and facilitate regulatory approval and commercialization of biotech crops. METHODS AND RESULTS: We report the overcome of the recalcitrance in sudangrass transformation and regeneration in order to use genome editing technique. Hence, an efficient regeneration system has been established to induce somatic embryogenesis from the immature inflorescence of two sudangrass cultivars on four MS-based media supplemented with different components. Our results indicate an interaction between genotype and medium composition. The combination of Giza-1 cultivar and M4 medium produces the maximum frequency of embryogenic calli of 80% and subsequent regeneration efficiency of 22.6%. Precise mutagenesis of the COMT gene is executed using the CRISPR/Cas9 system with the potential to reduce lignin content and enhance forage and biomass quality in sudangrass. CONCLUSION: A reliable regeneration and transformation system has been established for sudangrass using immature inflorescence, and the CRISPR/Cas9 system has demonstrated a promising technology for genome editing. The outcomes of this research will pave the road for further improvement of various sorghum genotypes to meet the global demand for food, feed, and biofuels, achieving sustainable development goals (SDGs).',
    'HIV envelope protein (Env) is the sole target of broadly neutralizing antibodies (BNAbs) that are capable of neutralizing diverse strains of HIV. While BNAbs develop spontaneously in a subset of HIV-infected patients, efforts to design an envelope protein-based immunogen to elicit broadly neutralizing antibody responses have so far been unsuccessful. It is hypothesized that a primary barrier to eliciting BNAbs is the fact that HIV envelope proteins bind poorly to the germline-encoded unmutated common ancestor (UCA) precursors to BNAbs. To identify variant forms of Env with increased affinities for the UCA forms of BNAbs 4E10 and 10E8, which target the Membrane Proximal External Region (MPER) of Env, libraries of randomly mutated Env variants were expressed in a yeast surface display system and screened using fluorescence activated cell sorting for cells displaying variants with enhanced abilities to bind the UCA antibodies. Based on analyses of individual clones obtained from the screen and on next-generation sequencing of sorted libraries, distinct but partially overlapping sets of amino acid substitutions conferring enhanced UCA antibody binding were identified. These were particularly enriched in substitutions of arginine for highly conserved tryptophan residues. The UCA-binding variants also generally exhibited enhanced binding to the mature forms of anti-MPER antibodies. Mapping of the identified substitutions into available structures of Env suggest that they may act by destabilizing both the initial pre-fusion conformation and the six-helix bundle involved in fusion of the viral and cell membranes, as well as providing new or expanded epitopes with increased accessibility for the UCA antibodies.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 512]

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

Training Details

Training Dataset

parquet

Dataset: parquet
Size: 33,870,508 training samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 3 tokens
mean: 22.56 tokens
max: 64 tokens
min: 12 tokens
mean: 250.53 tokens
max: 512 tokens

Samples:

anchor	positive
`Characteristics of the HIV/AIDS Epidemic among People Aged ≥ 50 Years in China during 2018-2021.`	Objective: This study aimed to determine the current epidemiological status of PLWHA aged ≥ 50 years in China from 2018 to 2021. It also aimed to recommend targeted interventions for the prevention and treatment of HIV/AIDS in elderly patients. Methods: Data on newly reported cases of PLWHA, aged ≥ 50 years in China from 2018 to 2021, were collected using the CRIMS. Trend tests and spatial analyses were also conducted. Results: Between 2018 and 2021, 237,724 HIV/AIDS cases were reported among patients aged ≥ 50 years in China. The main transmission route was heterosexual transmission (91.24%). Commercial heterosexual transmission (CHC) was the primary mode of transmission among males, while non-marital non-CHC ([NMNCHC]; 60.59%) was the prevalent route in women. The proportion of patients with CHC decreased over time ( Z = 67.716, P < 0.01), while that of patients with NMNCHC increased ( Z = 153.05, P < 0.01). The sex ratio varied among the different modes of infection, and it peaked a...
`Obstructive sleep apnea syndrome: A frequent and difficult-to-detect complication of radiotherapy for oropharyngeal cancers.`	`This pilot study reveals a higher prevalence of obstructive sleep apnea syndrome (OSAS) in patients treated for oropharyngeal squamous cell carcinoma with radiotherapy compared to the general population. OSAS indicators such as the Epworth Sleepiness Scale seem insufficient in the diagnostic approach to OSAS in this population and systematic screenings should be considered.`
`Two new JK silencing alleles identified by single molecule sequencing with 20-Kb long-reads.`	`BACKGROUND: The Kidd blood group gene SLC14A1 and JK[object Object]02 having c.743C>A did not express the Jkb and Jk3 antigens. CONCLUSIONS: We identified new JK silencing alleles and their critical SNVs by single-molecule sequencing and the findings were confirmed by transfection and expression studies.`

Loss: MultipleNegativesRankingLoss with these parameters:
json
1{ 2 "scale": 20.0, 3 "similarity_fct": "cos_sim" 4}

Evaluation Dataset

parquet

Dataset: parquet
Size: 33,870,508 evaluation samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 3 tokens
mean: 22.47 tokens
max: 95 tokens
min: 7 tokens
mean: 251.6 tokens
max: 512 tokens

Samples:

anchor	positive
`Causes and Management of Endoscopic Retrograde Cholangiopancreatography-Related Perforation: A Retrospective Study.`	BACKGROUND: Endoscopic retrograde cholangiopancreatography of ERCP-related perforation and conducted a retrospective review. RESULTS: Of the 15 patients, 6 were female and 9 were male, and the mean age was 77.1 years. According to Stapfer's classification, the 15 cases of ERCP-related perforation comprised 3 type I (duodenum), 3 type II (periampullary), 9 type III (distal bile duct or pancreatic duct), and no type IV cases. Fourteen of 15 (92.6%) were diagnosed during ERCP. The main cause of perforation was scope-induced damage, endoscopic sphincterotomy, and instrumentation penetration in type I, II, and III cases, respectively. Four patients with severe abdominal pain and extraluminal fluid collection underwent emergency surgery for repair and drainage. One type III patient with distal bile duct cancer underwent pancreaticoduodenectomy on day 6. Three type III patients with only retroperitoneal gas on computed tomography (CT) performed immediately after ERCP had no symptoms and neede...
`Covariance among premating, post-copulatory and viability fitness components in Drosophila melanogaster and their influence on paternity measurement.`	In polyandrous mating systems, male fitness depends on success in premating, post-copulatory and offspring viability episodes of selection. We tracked male success across all of these episodes simultaneously, using transgenic Drosophila melanogaster with ubiquitously expressed green fluorescent protein (that is GFP) in a series of competitive and noncompetitive matings. This approach permitted us to track paternity-specific viability over all life stages and to distinguish true competitive fertilization success from differential early offspring viability. Relationships between episodes of selection were generally not present when paternity was measured in eggs; however, positive correlations between sperm competitive success and offspring viability became significant when paternity was measured in adult offspring. Additionally, we found a significant male × female interaction on hatching success and a lack of repeatability of offspring viability across a focal male's matings, which may...
`Strategic partnerships to improve surgical care in the Asia–Pacific region: proceedings`	Emergency and essential surgery is a critical component of universal health coverage. Session three of the three-part virtual meeting series on Strategic Planning to Improve Surgical, Obstetric, Anaesthesia, and Trauma Care in the Asia–Pacific Region focused on strategic partnerships. During this session, a range of partner organisations, including intergovernmental organisations, professional associations, academic and research institutions, non-governmental organisations, and the private sector provided an update on their work in surgical system strengthening in the Asia–Pacific region. Partner organisations could provide technical and implementation support for National Surgical, Obstetric, and Anaesthesia Planning (NSOAP) in a number of areas, including workforce strengthening, capacity building, guideline development, monitoring and evaluation, and service delivery. Participants emphasised the importance of several forms of strategic collaboration: 1) collaboration across the spec...

Loss: MultipleNegativesRankingLoss with these parameters:
json
1{ 2 "scale": 20.0, 3 "similarity_fct": "cos_sim" 4}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 128
learning_rate: 2e-05
num_train_epochs: 1
max_steps: 251382
log_level: info
fp16: True
dataloader_num_workers: 16
load_best_model_at_end: True
resume_from_checkpoint: 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: 128
per_device_eval_batch_size: 8
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: 251382
lr_scheduler_type: linear
lr_scheduler_kwargs: {}
warmup_ratio: 0.0
warmup_steps: 0
log_level: info
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: 16
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: True
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

Click to expand

Epoch	Step	Training Loss	Validation Loss
0.0000	1	1.6269	-
0.0040	1000	0.2123	-
0.0080	2000	0.1191	-
0.0119	3000	0.0948	-
0.0159	4000	0.0824	-
0.0199	5000	0.0708	-
0.0239	6000	0.0665	-
0.0278	7000	0.0612	-
0.0318	8000	0.0578	-
0.0358	9000	0.0542	-
0.0398	10000	0.0528	-
0.0438	11000	0.0505	-
0.0477	12000	0.0461	-
0.0517	13000	0.0468	-
0.0557	14000	0.0442	-
0.0597	15000	0.0435	-
0.0636	16000	0.0414	-
0.0676	17000	0.0421	-
0.0716	18000	0.0399	-
0.0756	19000	0.0409	-
0.0796	20000	0.0393	-
0.0835	21000	0.0369	-
0.0875	22000	0.0349	-
0.0915	23000	0.0361	-
0.0955	24000	0.0358	-
0.0994	25000	0.0348	-
0.1034	26000	0.032	-
0.1074	27000	0.0341	-
0.1114	28000	0.0339	-
0.1154	29000	0.0325	-
0.1193	30000	0.0331	-
0.1233	31000	0.0306	-
0.1273	32000	0.0302	-
0.1313	33000	0.0304	-
0.1353	34000	0.0304	-
0.1392	35000	0.0306	-
0.1432	36000	0.0291	-
0.1472	37000	0.0273	-
0.1512	38000	0.0284	-
0.1551	39000	0.0292	-
0.1591	40000	0.0287	-
0.1631	41000	0.0277	-
0.1671	42000	0.0283	-
0.1711	43000	0.0268	-
0.1750	44000	0.027	-
0.1790	45000	0.0268	-
0.1830	46000	0.0259	-
0.1870	47000	0.0257	-
0.1909	48000	0.0252	-
0.1949	49000	0.0257	-
0.1989	50000	0.026	-
0.2029	51000	0.0262	-
0.2069	52000	0.0253	-
0.2108	53000	0.0252	-
0.2148	54000	0.025	-
0.2188	55000	0.0234	-
0.2228	56000	0.0233	-
0.2267	57000	0.0239	-
0.2307	58000	0.023	-
0.2347	59000	0.0246	-
0.2387	60000	0.0232	-
0.2427	61000	0.0244	-
0.2466	62000	0.0238	-
0.2506	63000	0.0231	-
0.2546	64000	0.0231	-
0.2586	65000	0.0226	-
0.2625	66000	0.0233	-
0.2665	67000	0.022	-
0.2705	68000	0.0222	-
0.2745	69000	0.0227	-
0.2785	70000	0.0232	-
0.2824	71000	0.0221	-
0.2864	72000	0.0223	-
0.2904	73000	0.0224	-
0.2944	74000	0.0218	-
0.2983	75000	0.0216	-
0.3023	76000	0.0213	-
0.3063	77000	0.0206	-
0.3103	78000	0.0214	-
0.3143	79000	0.0215	-
0.3182	80000	0.022	-
0.3222	81000	0.0209	-
0.3262	82000	0.0211	-
0.3302	83000	0.0215	-
0.3342	84000	0.0205	-
0.3381	85000	0.0201	-
0.3421	86000	0.0198	-
0.3461	87000	0.0208	-
0.3501	88000	0.0206	-
0.3540	89000	0.0193	-
0.3580	90000	0.0217	-
0.3620	91000	0.0197	-
0.3660	92000	0.0206	-
0.3700	93000	0.0193	-
0.3739	94000	0.019	-
0.3779	95000	0.0197	-
0.3819	96000	0.02	-
0.3859	97000	0.0176	-
0.3898	98000	0.0198	-
0.3938	99000	0.0186	-
0.3978	100000	0.0191	-
0.4018	101000	0.0187	-
0.4058	102000	0.0192	-
0.4097	103000	0.0183	-
0.4137	104000	0.0192	-
0.4177	105000	0.019	-
0.4217	106000	0.0179	-
0.4256	107000	0.0195	-
0.4296	108000	0.0183	-
0.4336	109000	0.018	-
0.4376	110000	0.0187	-
0.4416	111000	0.0178	-
0.4455	112000	0.0178	-
0.4495	113000	0.0181	-
0.4535	114000	0.0176	-
0.4575	115000	0.0189	-
0.4614	116000	0.0181	-
0.4654	117000	0.0185	-
0.4694	118000	0.0178	-
0.4734	119000	0.0183	-
0.4774	120000	0.0171	-
0.4813	121000	0.0164	-
0.4853	122000	0.0177	-
0.4893	123000	0.0184	-
0.4933	124000	0.0169	-
0.4972	125000	0.017	-
0.5012	126000	0.0174	-
0.5052	127000	0.0175	-
0.5092	128000	0.0167	-
0.5132	129000	0.0178	-
0.5171	130000	0.018	-
0.5211	131000	0.0175	-
0.5251	132000	0.0174	-
0.5291	133000	0.0176	-
0.5331	134000	0.0179	-
0.5370	135000	0.0171	-
0.5410	136000	0.0175	-
0.5450	137000	0.0175	-
0.5490	138000	0.0166	-
0.5529	139000	0.0168	-
0.5569	140000	0.0164	-
0.5609	141000	0.0163	-
0.5649	142000	0.0161	-
0.5689	143000	0.0169	-
0.5728	144000	0.0162	-
0.5768	145000	0.0171	-
0.5808	146000	0.0163	-
0.5848	147000	0.0163	-
0.5887	148000	0.0163	-
0.5927	149000	0.0164	-
0.5967	150000	0.0159	-
0.6007	151000	0.0164	-
0.6047	152000	0.0167	-
0.6086	153000	0.0167	-
0.6126	154000	0.0166	-
0.6166	155000	0.0157	-
0.6206	156000	0.0162	-
0.6245	157000	0.0164	-
0.6285	158000	0.0164	-
0.6325	159000	0.016	-
0.6365	160000	0.0162	-
0.6405	161000	0.0154	-
0.6444	162000	0.015	-
0.6484	163000	0.0158	-
0.6524	164000	0.0157	-
0.6564	165000	0.0165	-
0.6603	166000	0.0149	-
0.6643	167000	0.0159	-
0.6683	168000	0.0154	-
0.6723	169000	0.0156	-
0.6763	170000	0.0153	-
0.6802	171000	0.0155	-
0.6842	172000	0.0158	-
0.6882	173000	0.0144	-
0.6922	174000	0.0154	-
0.6961	175000	0.0153	-
0.7001	176000	0.0149	-
0.7041	177000	0.0152	-
0.7081	178000	0.0157	-
0.7121	179000	0.0148	-
0.7160	180000	0.0146	-
0.7200	181000	0.0152	-
0.7240	182000	0.0151	-
0.7280	183000	0.0159	-
0.7320	184000	0.0147	-
0.7359	185000	0.0139	-
0.7399	186000	0.0149	-
0.7439	187000	0.0143	-
0.7479	188000	0.0145	-
0.7518	189000	0.0154	-
0.7558	190000	0.0151	-
0.7598	191000	0.0155	-
0.7638	192000	0.016	-
0.7678	193000	0.0139	-
0.7717	194000	0.0154	-
0.7757	195000	0.0138	-
0.7797	196000	0.0147	-
0.7837	197000	0.0152	-
0.7876	198000	0.0141	-
0.7916	199000	0.0142	-
0.7956	200000	0.0149	-
0.7996	201000	0.0142	-
0.8036	202000	0.015	-
0.8075	203000	0.0142	-
0.8115	204000	0.0152	-
0.8155	205000	0.0142	-
0.8195	206000	0.0141	-
0.8234	207000	0.0146	-
0.8274	208000	0.014	-
0.8314	209000	0.0146	-
0.8354	210000	0.0138	-
0.8394	211000	0.0141	-
0.8433	212000	0.0143	-
0.8473	213000	0.0139	-
0.8513	214000	0.0138	-
0.8553	215000	0.0146	-
0.8592	216000	0.014	-
0.8632	217000	0.0138	-
0.8672	218000	0.0143	-
0.8712	219000	0.0151	-
0.8752	220000	0.0146	-
0.8791	221000	0.0135	-
0.8831	222000	0.0136	-
0.8871	223000	0.0139	-
0.8911	224000	0.0136	-
0.8950	225000	0.0142	-
0.8990	226000	0.0134	-
0.9030	227000	0.0143	-
0.9070	228000	0.0142	-
0.9110	229000	0.0142	-
0.9149	230000	0.0138	-
0.9189	231000	0.0136	-
0.9229	232000	0.0138	-
0.9269	233000	0.0144	-
0.9309	234000	0.0137	-
0.9348	235000	0.0135	-
0.9388	236000	0.014	-
0.9428	237000	0.014	-
0.9468	238000	0.0136	-
0.9507	239000	0.0134	-
0.9547	240000	0.0144	-
0.9587	241000	0.0136	-
0.9627	242000	0.014	-
0.9667	243000	0.0138	-
0.9706	244000	0.0133	-
0.9746	245000	0.0142	-
0.9786	246000	0.0135	-
0.9826	247000	0.013	-
0.9865	248000	0.0138	-
0.9905	249000	0.0146	-
0.9945	250000	0.0142	-
0.9985	251000	0.0134	-
1.0000	251382	-	0.0013

Framework Versions

Python: 3.11.11
Sentence Transformers: 3.4.1
Transformers: 4.48.2
PyTorch: 2.6.0+cu124
Accelerate: 1.3.0
Datasets: 3.2.0
Tokenizers: 0.21.0

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",
}

MultipleNegativesRankingLoss

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}
}

	anchor	positive
type	string	string
details	min: 3 tokens mean: 22.56 tokens max: 64 tokens	min: 12 tokens mean: 250.53 tokens max: 512 tokens