SentenceTransformer based on BAAI/bge-base-en-v1.5

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5. 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: BAAI/bge-base-en-v1.5
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 tokens
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': True}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, '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("aritrasen/bge-base-en-v1.5-ft")
# Run inference
sentences = [
    '# Serve and Deploy LLMs\n\nThis document shows how you can serve a LitGPT for deployment. \n\n&nbsp;\n## Serve an LLM\n\nThis section illustrates how we can set up an inference server for a phi-2 LLM using `litgpt serve` that is minimal and highly scalable.\n\n\n&nbsp;\n## Step 1: Start the inference server\n\n\n```bash\n# 1) Download a pretrained model (alternatively, use your own finetuned model)\nlitgpt download --repo_id microsoft/phi-2\n\n# 2) Start the server\nlitgpt serve --checkpoint_dir checkpoints/microsoft/phi-2\n```\n\n> [!TIP]\n> Use `litgpt serve --help` to display additional options, including the port, devices, LLM temperature setting, and more.\n\n\n&nbsp;\n## Step 2: Query the inference server\n\nYou can now send requests to the inference server you started in step 2. For example, in a new Python session, we can send requests to the inference server as follows:\n\n\n```python\nimport requests, json\n\nresponse = requests.post(\n    "http://127.0.0.1:8000/predict", \n    json={"prompt": "Fix typos in the following sentence: Exampel input"}\n)\n\nprint(response.json()["output"])\n```\n\nExecuting the code above prints the following output:\n\n```\nInstruct: Fix typos in the following sentence: Exampel input\nOutput: Example input.\n```',
    'How can I start an inference server for a phi-2 LLM using litgpt serve?\n',
    'What command do I use to convert the finetuned model to a HF transformer model?\n',
]
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]

Training Details

Training Dataset

Unnamed Dataset

Size: 21 training samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 51 tokens
mean: 424.62 tokens
max: 512 tokens

min: 12 tokens
mean: 17.19 tokens
max: 26 tokens

	anchor	positive
type	string	string
details	min: 51 tokens mean: 424.62 tokens max: 512 tokens	min: 12 tokens mean: 17.19 tokens max: 26 tokens

Samples:

anchor	positive
	7 B
1. Follow the instructions above to load the model into a Hugging Face transformers model. 2. Create a model.safetensor file: python<br>model.save_pretrained("out/hf-tinyllama/converted/")<br> 3. Copy the tokenizer files into the model-containing directory: bash<br>cp checkpoints/$repo_id/tokenizer* out/hf-tinyllama/converted<br> 4. Run the evaluation harness, for example: bash<br>lm_eval --model hf \<br> --model_args pretrained=out/hf-tinyllama/converted \<br> --tasks "hellaswag,gsm8k,truthfulqa_mc2,mmlu,winogrande,arc_challenge" \<br> --device "cuda:0" \<br> --batch_size 4<br>	`What is the command to run the evaluation harness?`
The LM Evaluation Harness requires a tokenizer to be present in the model checkpoint folder, which we can copy from the original download checkpoint: bash<br># Copy the tokenizer needed by the Eval Harness<br>cp checkpoints/microsoft/phi-2/tokenizer*<br>out/converted_model<br> Then, we can run the Evaluation Harness as follows: bash<br>lm_eval --model hf \<br> --model_args pretrained="out/converted_model" \<br> --tasks "hellaswag,gsm8k,truthfulqa_mc2,mmlu,winogrande,arc_challenge" \<br> --device "cuda:0" \<br> --batch_size 4<br> > [!TIP] > The Evaluation Harness tasks above are those used in Open LLM Leaderboard. You can find a list all supported tasks here. More information and additional resources - tutorials/convert_lit_models: Tutorial on converting LitGPT weights ## Get involved! We appreciate your feedback and contributions. If you have feature requests, questions, or want to contribute code or config files, please don't hesitate to use the GitHub Issue tracker. We welcome all individual contributors, regardless of their level of experience or hardware. Your contributions are valuable, and we are excited to see what you can accomplish in this collaborative and supportive environment. > [!TIP] > Unsure about contributing? Check out our How to Contribute to LitGPT guide. If you have general questions about building with LitGPT, please join our Discord.	`What is the command to run the Evaluation Harness?`

Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

Unnamed Dataset

Size: 10 evaluation samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 273 tokens
mean: 460.8 tokens
max: 512 tokens

min: 10 tokens
mean: 20.1 tokens
max: 34 tokens

	anchor	positive
type	string	string
details	min: 273 tokens mean: 460.8 tokens max: 512 tokens	min: 10 tokens mean: 20.1 tokens max: 34 tokens

Samples:

anchor	positive
(this table was sourced from the author's README) ## Download datasets You can download the data using git lfs: bash<br># Make sure you have git-lfs installed (https://git-lfs.com):<br>sudo apt install git-lfs<br> bash<br>git clone https://huggingface.co/datasets/cerebras/slimpajama-627b data/slimpajama-raw<br>git clone https://huggingface.co/datasets/bigcode/starcoderdata data/starcoderdata-raw<br> Around 1.2 TB of disk space is required to store both datasets. ## Prepare the datasets for training In order to start pretraining litgpt on it, you need to read, tokenize, and write the data in binary chunks. This will leverage the litdata optimization pipeline and streaming dataset. First, install additional dependencies for preprocessing: bash<br>pip install '.[all]'<br> You will need to have the tokenizer config available: bash<br>litgpt download \<br> --repo_id meta-llama/Llama-2-7b-hf \<br> --access_token your_hf_token \<br> --tokenizer_only true<br> Then, run the preprocessing script for each dataset and split. You will require 1.1 TB of disk space for Starcoder and 2.5 TB of space for the SlimPajama dataset. Starcoder: bash<br>python litgpt/data/prepare_starcoder.py \<br> --input_dir data/starcoderdata-raw \<br> --output_dir data/starcoder \<br> --tokenizer_path checkpoints/meta-llama/Llama-2-7b-hf<br> SlimPajama: bash<br>python litgpt/data/prepare_slimpajama.py \<br> --input_dir data/slimpajama-raw/validation \<br> --output_dir data/slimpajama/val \<br> --tokenizer_path checkpoints/meta-llama/Llama-2-7b-hf<br><br>python litgpt/data/prepare_slimpajama.py \<br> --input_dir data/slimpajama-raw/test \<br> --output_dir data/slimpajama/test \<br> --tokenizer_path checkpoints/meta-llama/Llama-2-7b-hf<br><br>python litgpt/data/prepare_slimpajama.py \<br> --input_dir data/slimpajama-raw/train \<br> --output_dir data/slimpajama/train \<br> --tokenizer_path checkpoints/meta-llama/Llama-2-7b-hf<br>	`How much disk space is required to store the SlimPajama dataset?`
# Serve and Deploy LLMs This document shows how you can serve a LitGPT for deployment. ## Serve an LLM This section illustrates how we can set up an inference server for a phi-2 LLM using litgpt serve that is minimal and highly scalable. ## Step 1: Start the inference server bash<br># 1) Download a pretrained model (alternatively, use your own finetuned model)<br>litgpt download --repo_id microsoft/phi-2<br><br># 2) Start the server<br>litgpt serve --checkpoint_dir checkpoints/microsoft/phi-2<br> > [!TIP] > Use litgpt serve --help to display additional options, including the port, devices, LLM temperature setting, and more. ## Step 2: Query the inference server You can now send requests to the inference server you started in step 2. For example, in a new Python session, we can send requests to the inference server as follows: python<br>import requests, json<br><br>response = requests.post(<br> "http://127.0.0.1:8000/predict", <br> json={"prompt": "Fix typos in the following sentence: Exampel input"}<br>)<br><br>print(response.json()["output"])<br> Executing the code above prints the following output: <br>Instruct: Fix typos in the following sentence: Exampel input<br>Output: Example input.<br>	`How can I start an inference server for a phi-2 LLM using litgpt serve?`
# TPU support This project utilizes Fabric, which supports TPUs via PyTorch XLA. > [!NOTE] > This guide assumes that you have already set-up your Google Cloud environment. To set up a Google Cloud instance with a TPU v4 VM, run the following commands: shell<br>gcloud compute tpus tpu-vm create litgpt --version=tpu-vm-v4-base --accelerator-type=v4-8 --zone=us-central2-b<br>gcloud compute tpus tpu-vm ssh litgpt --zone=us-central2-b<br> You can also choose a different TPU type. To do so, change the version, accelerator-type, and zone arguments. Find all regions and zones here. Multihost caveats TPU v4-8 uses a single host. SSH'ing into the machine and running commands manually will only work when using a single host (1 slice in the TPU pod). In multi-host environments, such as larger TPU pod slices, it's necessary to launch all commands on all hosts simultaneously to avoid hangs. For local development, it is advisable to upload a zip file containing all your current changes and execute it inside the VM from your personal computer: ```shell # Zip the local directory, excluding large directories from the zip. You may want to keep them. zip -r local_changes.zip . -x ".git/" "checkpoints/" "data/" "out/" # Copy the .zip file to the TPU VM gcloud compute tpus tpu-vm scp --worker=all local_changes.zip "litgpt:~" # Unzip on each host gcloud compute tpus tpu-vm ssh litgpt --worker=all --command="cd ~; unzip -q -o local_changes.zip" # Example of a typical workflow gcloud compute tpus tpu-vm ssh tmp --worker=all --command="cd ~; bash install_dependencies.sh" gcloud compute tpus tpu-vm ssh tmp --worker=all --command="cd ~; bash prepare_checkpoints.sh" gcloud compute tpus tpu-vm ssh tmp --worker=all --command="cd ~; bash run_desired_script.sh"	`How does this project support TPUs?`

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: 5
per_device_eval_batch_size: 5
num_train_epochs: 5
warmup_ratio: 0.1
fp16: 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: 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
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: 5
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
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	loss
0.4	2	0.6407	0.4190
0.8	4	0.7873	0.2789
1.2	6	0.1871	0.2089
1.6	8	0.2125	0.1718
2.0	10	0.0374	0.1648
2.4	12	0.1923	0.1695
2.8	14	0.0183	0.1723
3.2	16	0.1582	0.1770
3.6	18	0.0032	0.1824
4.0	20	0.0015	0.1870
4.4	22	0.1399	0.1901
4.8	24	0.002	0.1914

Framework Versions

Python: 3.10.12
Sentence Transformers: 3.0.1
Transformers: 4.41.2
PyTorch: 2.3.0+cu121
Accelerate: 0.27.0
Datasets: 2.20.0
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}
}