Ad: Nythos Embed Zero

README.md

@@ -6,15 +6,15 @@ pipeline_tag: sentence-similarity
 tags: [sentence-transformers, embedding, multilingual, turkish, retrieval, audio, image, norovox]
 ---
 
-# Mnemos
+# Nythos Embed Zero
 
-Mnemos is a multilingual embedding model from Norovox Labs, from our open-base line (its own-architecture counterpart is Norovox Embed Zero). It covers text, Turkish, retrieval, classification, clustering, audio and images, and is built on EmbeddingGemma 2. The text path has 271M parameters.
+Nythos Embed Zero is a multilingual embedding model from Norovox Labs, from the Nythos open-base line (its own-architecture counterpart is Norovox Embed Zero). It covers text, Turkish, retrieval, classification, clustering, audio and images, and is built on EmbeddingGemma 2. The text path has 271M parameters.
 
 > Every number below was measured by us under one protocol: fp32, the same prompts, and the same test sets for both models. Each test set was run once, after model selection on separate development sets.
 
 ## Results vs. EmbeddingGemma 2: ahead on 20 of 22 measures, tied on 1, behind on 1
 
-| Test | EmbeddingGemma 2 | Mnemos |
+| Test | EmbeddingGemma 2 | Nythos Embed Zero |
 |---|---|---|
 | STS-B (English similarity, Spearman) | .8749 | **.8865** |
 | STS-TR (Turkish similarity, Spearman) | .7863 | **.8091** |
@@ -33,7 +33,7 @@ Paired bootstrap per query, 95% interval of the difference:
 
 ### Classification (logistic regression on embeddings) and clustering (V-measure): ahead on 8 of 8
 
-| Task | EmbeddingGemma 2 | Mnemos |
+| Task | EmbeddingGemma 2 | Nythos Embed Zero |
 |---|---|---|
 | Banking77 | .9293 | **.9357** |
 | Emotion | .7009 | **.7190** |
@@ -56,7 +56,7 @@ The only measure where it is behind: image→text R@5 is .999 vs 1.000, i.e. one
 | NFCorpus | .3608 | **.3878** | .3432 | **.3793** |
 
 ### Against larger open embedding models (same five text tests, each with its own prompts; previous text mixture, α = 0.6)
-| Test | Mnemos (271M) | Qwen3-Embedding-0.6B (596M) | multilingual-e5-large-instruct (560M) | BGE-M3 (568M) |
+| Test | Nythos Embed Zero (271M) | Qwen3-Embedding-0.6B (596M) | multilingual-e5-large-instruct (560M) | BGE-M3 (568M) |
 |---|---|---|---|---|
 | STS-B | .8873 | **.9113** | .8838 | .8485 |
 | STS-TR | .8067 | .7501 | **.8296** | .7978 |
@@ -76,7 +76,7 @@ Kapsül is our inference layer on top of vLLM (pooling runner, bf16, length-awar
 
 ```python
 from vllm import LLM
-llm = LLM(model="emrevrg/Mnemos", runner="pooling", dtype="bfloat16")
+llm = LLM(model="emrevrg/Nythos-Embed-Zero", runner="pooling", dtype="bfloat16")
 docs = ["title: none | text: " + d for d in documents]   # same prompt as prompt_name="Document"
 vectors = [o.outputs.embedding for o in llm.embed(docs)]
 ```
@@ -95,7 +95,7 @@ What we tried and dropped: a 12-layer shallow document encoder distilled from th
 ## Usage
 ```python
 from sentence_transformers import SentenceTransformer
-m = SentenceTransformer("emrevrg/Mnemos")
+m = SentenceTransformer("emrevrg/Nythos-Embed-Zero")
 q = m.encode(["What is the longest river in Turkey?"], prompt_name="SearchQuery")
 d = m.encode(["The Kızılırmak, at 1,355 km, is the longest river in Turkey."], prompt_name="Document")
 print(q @ d.T)