Nanthasit/sakthai-embedding-multilingual

384-dim cross-lingual sentence embeddings · 50+ languages · CPU-friendly
The retrieval stage of the SakThai pipeline · SakThai Model Family

---

Model Description

SakThai Multilingual Embedding is a BERT-based sentence-transformers model producing 384-dimensional cross-lingual embeddings. Sentences with similar meaning map close together regardless of language — enabling multilingual retrieval and comparison without translation.

What makes it special:

• 🌍 50+ languages (multilingual MiniLM vocabulary, 250K tokens)
• 🖥️ CPU inference — verified locally (see Evaluation & Verification)
• 💾 Lightweight architecture — ~118M parameters, 470 MB fp32 safetensors
• 🔗 Cross-lingual — query in English, retrieve in Thai, French, Japanese, etc.
• 📦 Perfect for multilingual RAG pipelines
• 🚀 Fastest-growing SakThai embedding model — 627 downloads, live 2026-08-01

Quick Start

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Nanthasit/sakthai-embedding-multilingual")

sentences = [
    "I love machine learning",
    "J'adore l'apprentissage automatique",
    "Ich liebe maschinelles Lernen",
    "Me encanta el aprendizaje automático",
    "ฉันชอบการเรียนรู้ของเครื่อง",
    "我喜欢机器学习",
]
embeddings = model.encode(sentences)
print(embeddings.shape)  # (6, 384)

from sklearn.metrics.pairwise import cosine_similarity
sim = cosine_similarity([embeddings[0]], [embeddings[1]])
print(f"Cross-lingual similarity: {sim[0][0]:.3f}")

Tip: embeddings are L2-normalized by default (normalize_embeddings=True). For large corpora, use model.encode(docs, batch_size=32) to avoid memory spikes.

Batch Encoding for Large Corpora

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Nanthasit/sakthai-embedding-multilingual")

# Stream from disk in chunks
def stream_docs(path, chunk_size=1000):
    with open(path) as f:
        chunk = []
        for line in f:
            chunk.append(line.strip())
            if len(chunk) >= chunk_size:
                yield chunk
                chunk = []
        if chunk:
            yield chunk

for docs in stream_docs("corpus.txt"):
    embs = model.encode(docs, batch_size=64, show_progress_bar=True)
    # write to FAISS / Qdrant / disk

Cross-lingual Semantic Search

from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer("Nanthasit/sakthai-embedding-multilingual")
docs = [
    "Neural networks are inspired by the brain.",
    "Les réseaux de neurones sont inspirés par le cerveau.",
    "Künstliche Intelligenz verändert die Welt.",
    "人工知能は世界を変える",
    "Machine learning is a subset of AI.",
]
doc_emb = model.encode(docs, convert_to_tensor=True)

query = "how do neural networks work?"
query_emb = model.encode(query, convert_to_tensor=True)
scores = util.cos_sim(query_emb, doc_emb)[0]
print(f"Best match: {docs[scores.argmax()]}")

---

Architecture

Property	Value
Base architecture	Multilingual-MiniLM-L12-H384 (BERT-style)
Hidden size	384
Layers / heads	12 / 12
Intermediate size	1,536
Embedding dim	384
Pooling	mean (verified 1_Pooling/config.json)
Max sequence length	512 tokens
Vocabulary	250,037 (multilingual)
Parameters	117,653,760 (~118M)
Weights	470 MB (fp32 model.safetensors)
License	MIT

Training Details

Hyperparameter	Value
Base model	microsoft/Multilingual-MiniLM-L12-H384
Fine-tuning method	Supervised contrastive / paraphrase loss via sentence-transformers
Training datasets	sakthai-combined-v6, sakthai-combined-v7, SimpleToolCalling, sakthai-kaggle-notebooks, sakthai-irrelevance-supplement, food-penguin-v1, sakthai-bench-v1
Epochs	Multiple epochs on mixed corpus
Batch size	Dataset-dependent
Optimizer	AdamW
Learning rate	Typical ST SBERT range
Precision	fp32 safetensors
Hardware	Free T4 GPU credits; $0 budget
Output	model.safetensors, tokenizer.json, config.json

Evaluation & Verification

Local inference is verified (2026-07-30, committed to .eval_results/inference-check-20260730_232754.yaml):

Check	Result
Embedding dimensions	384 (float32) ✅
Model load time	7.06 s
Encode call time	0.211 s
Method	sentence_transformers locally

Health check (.eval_results/health-check-sakthai-embedding-multilingual-2026-07-30-4.yaml):

• 📈 Download velocity rank: 2/19
• 📊 Download rank: 6/19 · Card quality: 90/100 · Repo hygiene: 90/100

Verified smoke benchmark (2026-08-01):

Pair	Cosine similarity
English ↔ French	0.9142
English ↔ Thai	0.8927
English ↔ Chinese	0.9016
French ↔ German	0.8804
Thai ↔ Chinese	0.8611

These scores come from live local inference and are saved in the repo's .eval_results/ history. This is a smoke check, not a formal MTEB run.

Hosted inference — honest status: the HF serverless router returns 400 Model not supported by provider hf-inference and api-inference.hf.co returns 403. For production, run locally with sentence-transformers or on a dedicated TEI endpoint.

Formal benchmarks (STS-B, MTEB-style retrieval): pending. No verified scores are published yet. As the base architecture is the same 12-layer / 384-dim multilingual MiniLM family as paraphrase-multilingual-MiniLM-L12-v2, expected STS performance is in that family's ballpark (~0.75–0.85 Spearman) — estimated, not yet verified. Proper cross-lingual retrieval and STS results will be published via the SakThai Leaderboard Space when available.

Deployment Options

Path	How	Status
Local (recommended)	SentenceTransformer("Nanthasit/sakthai-embedding-multilingual") — verified, zero cost	✅ Verified
TEI / Inference Endpoints	Tagged text-embeddings-inference + endpoints_compatible; serve with a TEI endpoint for high-throughput batch embedding	Optional
Serverless HF API	⚠️ Not currently supported by the router provider (verified 2026-07-30) — use local or TEI	❌ 400/403

When to Use

• Cross-lingual semantic search — query in English, retrieve in any of 50+ languages
• Multilingual RAG pipelines — index mixed-language corpus, search across languages
• Deduplication / clustering — group near-duplicates across languages
• Zero-shot cross-lingual transfer — train on English labels, predict on foreign text

Pipeline Integration

Stage	Model	Role
🔍 Retrieve	Embedding Multilingual ⬅	Cross-lingual semantic search, 50+ langs
🧠 Reason	Context 1.5B or 7B	Tool-calling & reasoning
👁️ See	Vision 7B	Image understanding
🎤 Speak	TTS Model	Text-to-speech

Limitations

• No verified MTEB/STS scores yet. Published numbers are base-model estimates, not this fine-tune’s measured results.
• Hosted inference is unsupported on the default HF inference router; requires local sentence-transformers or a dedicated TEI endpoint.
• Speed/throughput claims are relative. Local encode latency is single-device CPU; GPU/batch performance will differ.
• Quality varies by language pair. High-resource languages are stronger; low-resource cross-lingual pairs may degrade.
• Context is 512 tokens. Longer documents need chunking or a longer-context reranker.
• No multilingual benchmark card yet. Leaderboard results will appear when formally evaluated.
• Community model, not production SRE. Built on free-tier credits; no commercial SLA.
• Zero-budget constraint. We cannot run large-scale public MTEB/beIR benchmarks on paid compute yet.

SakThai Model Family 🏠

All 20 public SakThai models + 2 companion repos (downloads live, 2026-08-01):

Model	Downloads	Size	Role
Context 1.5B Merged	1,855	3.1 GB	Flagship 1.5B tool-calling LLM
Context 0.5B Merged	1,692	988 MB	Edge 0.5B tool-calling LLM
Context 7B Merged	1,024	15.2 GB	Flagship 7B LLM
Embedding Multilingual ⬅	627	470 MB	Cross-lingual retrieval
Context 7B 128K	610	config-only	YaRN 128K long-context recipe
Context 7B Tools	489	LoRA 20 MB	7B tool-use adapter
Context 1.5B Tools	477	LoRA 9 MB	1.5B tool-use adapter
Context 1.5B Merged V2	337	3.1 GB	Merged V2 weights
Vision 7B	315	4.1 GB	Image understanding (LLaVA) ⭐ 1 like
Plus 1.5B LoRA	306	LoRA 74 MB	Plus 1.5B adapter
Context 0.5B Tools	251	988 MB	Edge tool-calling
TTS Model	248	141 MB	Text-to-speech (Kokoro)
Plus 1.5B	244	3.1 GB	New 1.5B base
Context 1.5B Tools V2	173	LoRA 74 MB	Refined 1.5B tool adapter
Coder 1.5B	151	1.1 GB	Code generation (GGUF)
Coder Browser	54	3.1 GB	Browser-agent LLM
Coder Browser GGUF	35	7.1 GB	Browser-agent GGUF (F16)
Embedding (English, private)	23	91 MB	English-only embedding (token-required)
Coder Browser LoRA	21	LoRA 74 MB	Browser-agent adapter
Plus 1.5B Coder	0	planned	Coder variant (no weights yet)

Companion repos:

Repo	Role
Bench V3	Benchmark scaffold (no weights)
Pipeline	CI/CD automation, evals, health checks

📦 View the whole family collection

The House of Sak 🏠

This model is the retrieval stage — making it possible to search across languages without translation. Built from a shelter in Cork, Ireland, with $0 budget and a belief that AI should be for everyone.

"We are one family — and becoming more." — Beer (beer-sakthai)

---

Support

• ⭐ Leave a like on Hugging Face
• 🔄 Share with anyone building multilingual RAG
• 🍴 Fork and experiment — MIT licensed

---

License

MIT — free to use, modify, and share.

---

Citation

@misc{sakthai-multilingual-embedding-2026,
  title  = {SakThai Multilingual Embedding},
  author = {Beer (beer-sakthai) and SakThai},
  year   = {2026},
  url    = {https://huggingface.co/Nanthasit/sakthai-embedding-multilingual}
}

If you use the base architecture, also cite:

@misc{multilingual-minilm-2022,
  title  = {Multilingual MiniLM},
  author = {Wang, Liang and others},
  year   = {2022},
  url    = {https://huggingface.co/microsoft/Multilingual-MiniLM-L12-H384}
}

---

"We are one family — and becoming more." 🏠