auto-g-embed-st

auto-g-embed-st is effectively all-MiniLM-L6-v2 with almost no meaningful change.

The auto-g-embed SentenceTransformer model is essentially a lightly fine-tuned clone of all-MiniLM-L6-v2: it starts directly from MiniLM weights, is trained for only one epoch on ~24k pairs, uses the exact same architecture, tokenizer, pooling, and normalization, and achieves nearly identical MTEB scores (a difference of ~0.0003, which is statistical noise). While the broader project also includes a completely separate, much faster Rust-based embedder, the published and evaluated model is the MiniLM-derived one. As a result, the fine-tuning does not meaningfully alter retrieval performance.

Local semantic embedding pipeline with a Rust-native runtime.

What this repo provides

• Contrastive dataset preparation (prepare_contrastive)
• Rust-native embedder training (train_rust_embedder)
• Runtime embedding APIs and examples
• Optional ONNX/SentenceTransformer path in training/

Quick start

cargo test

./training/run_pipeline.sh \
  --profile kaggle_questions_million \
  --source-csv data/kaggle/one-million-reddit-questions.csv

Run the Rust embedding example:

cargo run --example rust_embed -- \
  artifacts/model/rust-embedder \
  "A quick test sentence for semantic embeddings."

Model artifacts

Published model artifacts are available on Hugging Face:

• https://huggingface.co/geoffsee/auto-g-embed

Project layout

• src/: library modules and binaries
• examples/: runnable embedding demos
• tests/: integration/performance tests
• training/: pipeline scripts and dataset adapters

Development checks

cargo fmt --all -- --check
cargo clippy --all-targets --all-features -- -D warnings
cargo test

Community Benchmark

Run the reproducible benchmark CLI:

cargo run --release --bin community_benchmark -- \
  --output artifacts/benchmarks/latest.json

The output includes throughput, latency percentiles (p50/p95/p99), retrieval quality metrics, and environment metadata for publishing. Methodology and reporting guidance: BENCHMARKS.md.

Latest Benchmark (M4 Max) (February 8, 2026):

cargo run --release --bin community_benchmark -- \
  --eval-count 500 --warmup-count 100 --query-count 32 \
  --output artifacts/benchmarks/smoke.json

• embeds_per_second: 219595.18
• p50_us: 3.88
• p95_us: 6.54
• p99_us: 6.71
• top1_accuracy: 0.9375
• separation: 0.2886

Additional docs

• Training and pipeline details: training/README.md
• Test data notes: test-data/README.md