bge-small-rrf-v4

A 33M-parameter (384-dim) English embedding model, fine-tuned from BAAI/bge-small-en-v1.5 using vstash's self-supervised hybrid-retrieval disagreement signal.

Reproduces the v3 training recipe on the post-v0.34 vstash pipeline. Same size, same speed, same recipe; trained against the corrected cosine distance metric instead of the buggy L2-as-cosine that pre-v0.34 applied. Use this as the "current code" reference. v3 remains valid as the pre-v0.34 reference snapshot.

Why v4 exists

The v0.34 release (vstash 0.34.0) fixed a latent bug in the vec_chunks table: sqlite-vec's vec0(embedding float[N]) defaults to L2 distance, but every comment, threshold, and telemetry field in vstash labelled the value "cosine distance." Worked accidentally for unit-normalized BGE because L2 and cosine are monotonically related, but the distance_cutoff and relevance_tier thresholds were calibrated against L2 ratios while the documentation and downstream code assumed cosine. v0.34 rescaled both (1.15 -> 1.3225 for the default cutoff; 0.95/0.98 -> 0.4513/0.4802 for the tier thresholds) to match the squared cosine convention.

bge-small-rrf-v3 was trained before that fix landed. v4 retrains the same recipe (arm_vol from the H-R9 ablation: temperature=0.5, total_triples=60000, MNRL, 2 epochs at lr=3e-6, batch=64, seed=42) on the post-v0.34 vstash code so the training mining runs through the corrected pipeline.

Empirically the difference is small. On the 5-dataset BEIR macro v4 vs v3 lands at delta = -0.0057 NDCG@10 -- within the calibration band. Per-dataset, v3 wins on the 3 training datasets (scifact, nfcorpus, fiqa) and v4 wins on the held-out 2 (scidocs, arguana), consistent with v3 having mildly overfit to a metric quirk that v4's training data does not contain. Reproduction notebook is at experiments/retrain_v4_post_v034_validation.ipynb in the vstash repo; the pre-committed decision rule applied during the ablation is at experiments/results/asymmetry_decision_2026_04_28.md.

Eval numbers

vstash hybrid pipeline (adaptive RRF + FTS5 + MMR dedup), v0.35 code, sentence-transformers backend.

Dataset	base BGE-small	bge-small-rrf-v4	Delta
SciFact	0.7251	0.7575	+0.0324
NFCorpus	0.3591	0.3707	+0.0116
FiQA	0.3917	0.4455	+0.0538
SciDocs	0.1945	0.1971	+0.0026
ArguAna	0.4367	0.4386	+0.0019
macro	0.4214	0.4419	+0.0205 (+4.9%)

Comparison notes:

• Base BGE-small here is BAAI/bge-small-en-v1.5 evaluated through the same vstash pipeline. Numbers above are the Colab T4 ST-CUDA measurement of v4. A local Mac ST-CPU re-measurement is pending and will be posted as a sidecar; the FastEmbed (ONNX) backend on current code is ~+0.003 NDCG@10 vs ST.
• vstash hybrid + base BGE on the v0.35 pipeline already exceeds ColBERTv2 published macro (0.4214 vs 0.402 = +4.8%). The fine-tune adds another +4.9% on top. Pipeline contribution dominates; the fine-tune is a modest edge for domain-specific use.
• v4 vs ColBERTv2 published per dataset: SciFact +9.3%, NFCorpus +7.8%, FiQA +25.1%, SciDocs +28.0%, ArguAna -5.3%.

When to use v4

• General BEIR-style retrieval where you want the post-v0.34 reference weights. Recommended default for new vstash deployments.
• Reproducing the paper v2 benchmark numbers under current code.
• As a baseline for further fine-tuning via vstash retrain.

When NOT to use v4

• Chat-memory retrieval. Use Stffens/bge-small-rrf-lme-v1 instead, which is trained on real labeled LongMemEval queries.
• Pre-v0.34 reproducibility. If you are reproducing measurements taken on a vstash version before 0.34.0, use Stffens/bge-small-rrf-v3 to match the training-time pipeline.

Usage

vstash

pip install vstash
vstash reindex --model Stffens/bge-small-rrf-v4

sentence-transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Stffens/bge-small-rrf-v4")
embs = model.encode(["query text", "document passage"], normalize_embeddings=True)

Training

• Base: BAAI/bge-small-en-v1.5 (33M params, 384 dim).
• Loss: MultipleNegativesRankingLoss (MNRL).
• Sampling: temperature=0.5 across SciFact, NFCorpus, FiQA.
• Total triples: 60,000 target budget (39,852 actual after mining).
• Per-dataset triples: SciFact 4291, NFCorpus 9713, FiQA 25848.
• Epochs: 2.
• LR: 3e-6.
• Batch size: 64.
• AMP: enabled.
• Seed: 42.
• Hardware: Colab T4 (CUDA), --bulk-mine --bulk-mine-device cuda.
• Wall time: ~9 minutes training.

Triples are mined inside vstash retrain from vector-vs-FTS top-K disagreement. The training distribution is dominated by FiQA (~65%) because FiQA queries trigger more disagreement events than the more homogeneous claim-style SciFact and NFCorpus corpora.

Citation

If you use this model, please cite the vstash v2 paper (in preparation). v1 is on arXiv:

@misc{steffens2026vstash,
  title  = {vstash: Local-First Hybrid Retrieval with Adaptive Fusion for LLM Agents},
  author = {Steffens, Jayson},
  year   = {2026},
  eprint = {2604.15484},
  archivePrefix = {arXiv}
}

License

MIT, same as BAAI/bge-small-en-v1.5.