Aizen Vault Embeddings

A fine-tuned sentence-transformers/all-MiniLM-L6-v2, trained to improve semantic search over a personal-assistant-style memory vault. See the companion aizen-vault-search Space (live demo) and aizen-protocol dataset (governing methodology) on this profile.

Training data — read this before drawing conclusions

Trained on 42 pairs built entirely from synthetic, fictional notes (fictional companies, a fictional country called "Velmoria," fictional government/financial institutions) written specifically for this project. No real personal or business data was used to train this checkpoint.

Honest eval — including a negative result

This model uses the same recipe (MultipleNegativesRankingLoss, hand-written natural-language query pairs + auto-generated description-to-body pairs) that produced a large, genuine improvement on a real private vault in earlier testing — there, the base model's top-1 result on a held-out query was wrong (country-canada for a question about a monarch's constitutional role) and the fine-tuned model corrected it (score 0.268 → 0.559 on the correct note). That test used real data and is not published, for confidentiality reasons.

Re-running the identical recipe on this synthetic dataset, evaluated on two held-out queries never seen in training:

Held-out query	Base top-1	Fine-tuned top-1
"which energy company has the fictional grid regulator been working with"	correct (same note)	correct (same note, same rank)
"how does the monarch's constitutional role work in this fictional nation"	correct (same note)	correct (same note, same rank)

The base model already got both right here. The synthetic notes' descriptions were written in language closer to natural questions than the real vault's terser, third-person style, so this particular test wasn't hard enough to show the same effect the real data showed. The fine-tune produced only minor, mixed changes in score margins and lower-rank ordering here — not the dramatic correction seen on real data.

What this means honestly: the technique is validated (on real data, not published here); this specific public checkpoint is a safe demonstration of the method, not a repeat of that result. Treat this checkpoint as a documented starting point, not a finished capability claim.

This is a sentence-transformers model finetuned from sentence-transformers/all-MiniLM-L6-v2. It maps inputs to a 384-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: sentence-transformers/all-MiniLM-L6-v2
Maximum Sequence Length: 256 tokens
Output Dimensionality: 384 dimensions
Similarity Function: Cosine Similarity
Supported Modality: Text

Model Sources

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

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'architecture': 'BertModel'})
  (1): Pooling({'embedding_dimension': 384, 'pooling_mode': 'mean', 'include_prompt': True})
  (2): Normalize({'module_input_name': 'sentence_embedding', 'module_output_name': 'sentence_embedding'})
)

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("sentence_transformers_model_id")
# Run inference
sentences = [
    'when is the annual report due and who audits it',
    "syn-annual-report-deadline\nSYNTHETIC EXAMPLE — the fictional company's annual report deadline and which external auditor signs off\nThe annual report is due to the fictional regulator by March 31st each year; the external auditor (a fictional firm, Kestrel & Boyd) needs the draft financials four weeks earlier to complete sign-off in time.\n\n**Why this matters**: a recurring compliance deadline with a real external dependency (the auditor's own lead time), the kind of fact that's costly to forget.",
    'example-quarterly-board-deck-deadline\nSYNTHETIC EXAMPLE — a recurring quarterly board-deck deadline and who owns which section\nThe quarterly board deck is due the first Monday of the new quarter. Finance owns the numbers section, product owns the roadmap slide, and the fictional CEO persona in this example always wants the risks slide reviewed by legal before it goes to the board, not after.\n\n**Why this matters**: recurring institutional deadlines with named owners are exactly what should live in permanent memory rather than being re-explained every quarter.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.6087, 0.2622],
#         [0.6087, 1.0000, 0.3761],
#         [0.2622, 0.3761, 1.0000]])

Training Details

Training Dataset

Unnamed Dataset

Size: 42 training samples
Columns: anchor and positive
Approximate statistics based on the first 42 samples:
anchor positive
type string string

modality text text
details
min: 5 tokens
mean: 20.21 tokens
max: 41 tokens

min: 90 tokens
mean: 118.95 tokens
max: 153 tokens

	anchor	positive
type	string	string
modality	text	text
details	min: 5 tokens mean: 20.21 tokens max: 41 tokens	min: 90 tokens mean: 118.95 tokens max: 153 tokens

Samples:

anchor	positive
`which deal has gone quiet`	example-solstice-energy-followup SYNTHETIC EXAMPLE — stalled outreach to Solstice Energy Partners, a fictional renewable-energy utility, flagged for a 30-day-silent follow-up Solstice Energy Partners (fictional entity) has gone quiet for over a month after an initially warm first call. Last real contact was a scoping email that was never answered. Rather than guessing at a reason, the honest status is: dormant, needs a low-pressure check-in, not a hard sell. Why this matters: a good follow-up digest should say "30+ days silent" plainly, not paper over it with an optimistic status that isn't earned by any real recent contact.
`any dietary restrictions I should know about for an event`	example-team-offsite-logistics SYNTHETIC EXAMPLE — logistics and dietary constraints for a team offsite, including a real constraint that was almost missed Planning a team offsite: venue booked, but one attendee has a shellfish allergy that almost didn't make it into the catering brief because it was mentioned once, verbally, three weeks earlier. Caught it by re-checking the original conversation rather than relying on a fresh summary. Why this matters: this is the practical case for a permanent, searchable memory over a chat history that scrolls away — the fact was real and important, it just wasn't recent.
`what does the AI governance researcher think about smaller regulators`	example-dr-chandran-research-call SYNTHETIC EXAMPLE — a research call with a fictional academic, Dr. Priya Chandran, on AI governance frameworks for emerging markets Spoke with Dr. Priya Chandran (fictional persona), a researcher working on AI governance in emerging-market contexts. She pointed to a real distinction worth remembering generally: policy frameworks written for large, well-resourced regulators often assume enforcement capacity that smaller regulatory bodies simply don't have — a good assistant should flag that gap rather than treat every jurisdiction's framework as equally enforceable on paper. Why this matters: this is the kind of nuance that's easy to flatten into a generic summary, and worth preserving instead.

Loss: MultipleNegativesRankingLoss with these parameters:

{
    "scale": 20.0,
    "similarity_fct": "cos_sim",
    "gather_across_devices": false,
    "directions": [
        "query_to_doc"
    ],
    "partition_mode": "joint",
    "hardness_mode": null,
    "hardness_strength": 0.0
}

Training Hyperparameters

Non-Default Hyperparameters

num_train_epochs: 15
warmup_steps: 0.1

All Hyperparameters

Click to expand

per_device_train_batch_size: 8
num_train_epochs: 15
max_steps: -1
learning_rate: 5e-05
lr_scheduler_type: linear
lr_scheduler_kwargs: None
warmup_steps: 0.1
optim: adamw_torch_fused
optim_args: None
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
optim_target_modules: None
gradient_accumulation_steps: 1
average_tokens_across_devices: True
max_grad_norm: 1.0
label_smoothing_factor: 0.0
bf16: False
fp16: False
bf16_full_eval: False
fp16_full_eval: False
tf32: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
use_liger_kernel: False
liger_kernel_config: None
use_cache: False
neftune_noise_alpha: None
torch_empty_cache_steps: None
auto_find_batch_size: False
log_on_each_node: True
logging_nan_inf_filter: True
include_num_input_tokens_seen: no
log_level: passive
log_level_replica: warning
disable_tqdm: False
project: huggingface
trackio_space_id: None
trackio_bucket_id: None
trackio_static_space_id: None
per_device_eval_batch_size: 8
prediction_loss_only: True
eval_on_start: False
eval_do_concat_batches: True
eval_use_gather_object: False
eval_accumulation_steps: None
include_for_metrics: []
batch_eval_metrics: False
save_only_model: False
save_on_each_node: False
enable_jit_checkpoint: False
push_to_hub: False
hub_private_repo: None
hub_model_id: None
hub_strategy: every_save
hub_always_push: False
hub_revision: None
load_best_model_at_end: False
ignore_data_skip: False
restore_callback_states_from_checkpoint: False
full_determinism: False
seed: 42
data_seed: None
use_cpu: False
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
parallelism_config: None
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_pin_memory: True
dataloader_persistent_workers: False
dataloader_prefetch_factor: None
dataloader_multiprocessing_context: None
dataloader_in_order: True
remove_unused_columns: True
label_names: None
train_sampling_strategy: random
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
ddp_static_graph: None
ddp_backend: None
ddp_timeout: 1800
fsdp: None
fsdp_config: None
deepspeed: None
debug: []
skip_memory_metrics: True
do_predict: False
resume_from_checkpoint: None
local_rank: -1
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}
warmup_ratio: None

Training Logs

Epoch	Step	Training Loss
0.8333	5	0.2045
1.6667	10	0.2811
2.5	15	0.1620
3.3333	20	0.1210
4.1667	25	0.1075
5.0	30	0.0356
5.8333	35	0.0401
6.6667	40	0.0379
7.5	45	0.1162
8.3333	50	0.1207
9.1667	55	0.1511
10.0	60	0.0806
10.8333	65	0.0953
11.6667	70	0.1007
12.5	75	0.1483
13.3333	80	0.1172
14.1667	85	0.2504
15.0	90	0.2687

Training Time

Training: 1.4 minutes

Framework Versions

Python: 3.13.14
Sentence Transformers: 6.1.0
Transformers: 5.17.0
PyTorch: 2.14.0+cpu
Accelerate: 1.15.0
Datasets: 5.0.1
Tokenizers: 0.23.2

Additional Resources

Training and Finetuning Embedding Models with Sentence Transformers: the end-to-end guide for training or finetuning Sentence Transformer models.
Introduction to Matryoshka Embedding Models: variable-size embeddings that can be truncated with minimal quality loss.
Binary and Scalar Embedding Quantization for Significantly Faster & Cheaper Retrieval: post-training compression of embedding vectors.
Multimodal Embedding & Reranker Models with Sentence Transformers: use text, image, audio, and video models through the same API.
Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers: train multimodal embedding models, with a Visual Document Retrieval walkthrough.

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{oord2019representationlearningcontrastivepredictive,
      title={Representation Learning with Contrastive Predictive Coding},
      author={Aaron van den Oord and Yazhe Li and Oriol Vinyals},
      year={2019},
      eprint={1807.03748},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/1807.03748},
}

Downloads last month: 18

Model tree for Marqjeev/aizen-vault-embeddings

Base model

nreimers/MiniLM-L6-H384-uncased

Quantized

sentence-transformers/all-MiniLM-L6-v2

Finetuned

(1128)

this model

Papers for Marqjeev/aizen-vault-embeddings

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Paper • 1908.10084 • Published Aug 27, 2019 • 18

Representation Learning with Contrastive Predictive Coding

Paper • 1807.03748 • Published Jul 10, 2018 • 1