GoktugD/DUSUNEN-Pusula-118M-v0

DUSUNEN Pusula 118M v0

A 117.7M-parameter, 384-dimensional Turkish retrieval experiment built to measure the quality, index-memory and browser-deployment trade-off of a smaller encoder. It is a retriever, not a chat or text-generation model.

Five-task measured results

Official MTEB 2.18.16 evaluators, Turkish-only subsets, normalized embeddings and explicit prompt formats were used for every row.

Model	Params	Dim	TurHist	XQuAD	WebFAQ	MKQA	Belebele	Macro
multilingual E5 base	278.0M	768	0.49726	0.95335	0.65032	0.07213	0.92503	0.619618
DUSUNEN Rota 270M v2	268.1M	640	0.42198	0.86393	0.56886	0.10331	0.88493	0.568602
DUSUNEN Rota 270M v1	268.1M	640	0.42196	0.85832	0.56402	0.10296	0.88222	0.565896
DUSUNEN Pusula 118M v0	117.7M	384	0.25299	0.81123	0.46307	0.04855	0.82451	0.480070

Relative to DUSUNEN Rota 270M v1, this model uses 56.1% fewer parameters and 40% smaller float32 vectors, with a 15.2% relative macro-score reduction. The source-domain triplet accuracy rose from 0.8745 to 0.9200. This is a measured size/quality option, not a universal quality replacement.

Use

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("GoktugD/DUSUNEN-Pusula-118M-v0")
query = model.encode("Hard negative neden önemlidir?", normalize_embeddings=True)
docs = model.encode(
    ["Zor negatifler karar sınırını güçlendirir.", "Ankara Türkiye'nin başkentidir."],
    normalize_embeddings=True,
)
print(docs @ query)

Queries and documents are encoded as plain text. Store normalized vectors and use cosine similarity or inner product.

Training and deployment

• Pinned base revision: e8f8c211226b894fcb81acc59f3b34ba3efd5f42
• Data: 100,000 train / 2,000 validation triplets
• Objective: Cached Multiple Negatives Ranking Loss
• Sequence length: 256; effective batch: 64; BF16; seed: 3407
• Hardware: one RTX 5060 Laptop GPU with 8 GB VRAM
• Training time: 1,397.5 seconds; peak allocated GPU memory: 1.073 GB

The included dynamic-int8 ONNX graph is 118,335,516 bytes. On 16 documents and four probes, PyTorch-to-INT8 mean cosine agreement was 0.990002, minimum agreement was 0.977650 and top-1 agreement was 100%.

Limitations

• The training data is machine-translated Turkish MS MARCO and inherits its translation artifacts, dated facts and web-text biases.
• Compactness reduces retrieval quality on this five-task suite.
• Similarity is a ranking signal, not a calibrated probability or fact check.
• Exact-overlap auditing does not detect paraphrases or semantic duplicates.

The release includes raw MTEB results, training state, environment metadata, ONNX validation and SHA-256 file manifests.