EIDORA/MiniLM_L6_v2_Text

MiniLM L6 v2 — Sentence Embeddings

MiniLM L6 v2 is a compact sentence-embedding model. EIDORA converts sentences into normalized 384-dimensional vectors for semantic comparison.

Training Data and Checkpoint Provenance

• Architecture: Six-layer MiniLM Transformer with sentence-level mean pooling.
• Original checkpoint: sentence-transformers/all-MiniLM-L6-v2.
• Pinned revision: 1110a243fdf4706b3f48f1d95db1a4f5529b4d41.
• Training data: Sentence-pair training data described by the upstream Sentence Transformers model card.
• Training method: Sentence-pair training described in the upstream Sentence Transformers model card.
• Source: all-MiniLM-L6-v2 (original author attribution retained).

Model Overview

• Model: MiniLM L6 v2 — Sentence Embeddings
• EIDORA identifier: eidora/text-minilm-l6-v2-light
• Modality: Text
• Embedding dimensions: 384
• Embedding normalization: L2-normalized
• Similarity measure: Cosine similarity
• Package schema: Version 3
• Model version: 1.0.0

EIDORA Use and Supported Modalities

Supported modality: Text only.

Researchers can use this package to:

• Compare short sentences by semantic similarity.
• Build simple text retrieval or search workflows.
• Cluster or explore groups of related passages.

Input and Preprocessing

EIDORA reads UTF-8 text and uses its packaged tokenizer from tokenizer/.

1. Add model-specific special tokens.
2. Pad or truncate to 256 tokens, using right-side padding and right-side truncation.
3. Pass integer input_ids (sourced from token_ids) and attention_mask tensors to ONNX.

Input tensor shapes: [batch, 256] for each input, dtype int64. The tokenizer and its configuration are included in the repository. Text beyond the configured maximum is truncated; callers should chunk longer inputs when needed.

Output Embeddings

The ONNX output is float32 with shape ['batch', 384]. Each vector has 384 dimensions, is L2-normalized, and is intended for cosine similarity comparisons.

EIDORA encodes text using the packaged tokenizer and performs attention-mask-aware mean pooling with L2 normalization.

Runtime and Compatibility

The package contains model.onnx, its tokenizer files and schema-v3 text input preparation.

It declares CPUExecutionProvider. Actual execution availability depends on the installed ONNX Runtime and EIDORA Desktop environment. config.yaml is authoritative for graph inputs, preprocessing, provider and embedding dimensions.

Limitations

• Long text is truncated at 256 tokens unless it is split into chunks.
• Similarity is not a measurement of factual correctness or truth.
• The original training distribution can differ from a specialized collection.

Results should be evaluated against the intended research data and questions.

Research Context and References

Main paper

Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

Paper: https://aclanthology.org/D19-1410/

The publication describes the upstream learning method and original model. The embeddings returned by EIDORA additionally depend on the specific preprocessing and output adaptation documented here.

Primary citation (BibTeX)

@inproceedings{reimers-2019-sentence-bert,
  title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
  author = "Reimers, Nils and Gurevych, Iryna",
  booktitle = "Proceedings of EMNLP-IJCNLP",
  year = "2019",
  pages = "3982--3992",
  url = "https://aclanthology.org/D19-1410/"
}

License and Attribution

The upstream sentence-transformers/all-MiniLM-L6-v2 checkpoint is distributed under Apache License 2.0. See the included LICENSE. EIDORA credits the Sentence Transformers contributors and provides the ONNX adaptation.

See the LICENSE file included in this repository.

Package Information

• Model repository: EIDORA/MiniLM_L6_v2_Text
• Package version: 1.0.0
• Package schema: 3
• EIDORA configuration: config.yaml
• ONNX graph(s): model.onnx
• Supported modality: Text
• License identifier: apache-2.0
• Declared provider: CPUExecutionProvider
• Required supplementary files: tokenizer/ (packaged tokenizer), LICENSE

The configuration defines the authoritative model-facing runtime contract. Other packaged assets, where present, do not introduce additional inference modalities.