Echo-DSRN-v0.1.3-Embed-Intent

98M-parameter multilingual intent classification embedding model based on the Echo-DSRN architecture (Dual-State Recurrent Neural Network) ◦ Recurrent Hybrid.

Fine-tuned from ethicalabs/Echo-DSRN-v0.1.3-Embed-Exp on Amazon MASSIVE across all 51 languages using MultipleNegativesRankingLoss (MNRL).

Model specs

Property	Value
Architecture	Echo-DSRN (Recurrent Neural Network)
Parameters	98,266,629 (~98M)
Layers	8 DSRN blocks
Hidden dim	512
Attention heads	4
Vocab size	32,017 tokens
Precision	fp32
Base model	ethicalabs/Echo-DSRN-v0.1.3-Embed-Exp
GPU	AMD Radeon AI Pro R9700 (ROCm 7.2)

MTEB Results

MassiveIntentClassification (60 intents, 51 languages)

Metric	Score
Accuracy (mean)	72.42%
Accuracy (min)	62.98%
Accuracy (max)	78.33%
F1 (mean)	66.27%
F1 (min)	56.52%
F1 (max)	71.94%

MassiveScenarioClassification (17 scenarios, 51 languages)

Metric	Score
Accuracy (mean)	79.00%
Accuracy (min)	71.62%
Accuracy (max)	84.30%
F1 (mean)	78.28%
F1 (min)	70.05%
F1 (max)	84.08%

Evaluated via MTEB v2.12.30 logistic regression protocol on frozen embeddings. Per-language scores available in the model-index metadata.

Training

• Base model: ethicalabs/Echo-DSRN-v0.1.3-Embed-Exp (STS-pretrained, 0.753 avg Spearman on MTEB STS)
• Dataset: Amazon MASSIVE, all 51 locales (~1M training utterances)
• Loss: MultipleNegativesRankingLoss with intent-grouped positive pairs
• Pooling: mean_c_all (2048-dim recurrent slow state)
• Convergence: Early stopping at epoch 1.2; linear accuracy gain (+2 pts/1k steps), no grokking plateau
• Random baseline: ~1.7% (60-class 1-NN)

Example Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("ethicalabs/Echo-DSRN-v0.1.3-Embed-Intent", trust_remote_code=True, device="cpu")

sentences = [
  "Can I order a pizza?",
  "I am so hungry. what about pizza?",
  "I like spaghetti."
]
embeddings = model.encode(sentences)

similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)

Console output

torch.Size([3, 3])
>>> print(similarities)
tensor([[1.0000, 0.9333, 0.7381],
        [0.9333, 1.0000, 0.8411],
        [0.7381, 0.8411, 1.0000]])

Intent Vector Space Behavior

Here is what the model is actually doing under the hood for each pair:

1. Sim(0, 1) = 0.9333 — Matching Actionable Intent

   • Sentence 0: "Can I order a pizza?"
   • Sentence 1: "I am so hungry. what about pizza?"
   • Analysis: Despite using completely different phrasing and syntax (one is a direct question, the other is a multi-sentence conversational prompt), the model maps them to nearly the same spot in vector space. The recurrent slow state identifies the underlying action (order_food) and topic (pizza), yielding a massive 0.9333 correlation.

2. Sim(0, 2) = 0.7381 — Action vs. Statement Separation

   • Sentence 0: "Can I order a pizza?"
   • Sentence 2: "I like spaghetti."
   • Analysis: Notice the significant drop down to 0.7381. Even though both sentences live in the general domain of Italian food, the model correctly separates an actionable transactional request ("Can I order...") from a static statement of personal preference ("I like..."). This is where the fine-tuning on the MASSIVE dataset with MultipleNegativesRankingLoss shines: it prevents the model from relying purely on lexical topic overlap.

3. Sim(1, 2) = 0.8411 — Conversational Context

   • Sentence 1: "I am so hungry. what about pizza?"
   • Sentence 2: "I like spaghetti."
   • Analysis: This pair scores higher (0.8411) than (0, 2). Because Sentence 1 expresses a state/desire ("I am so hungry"), its semantic profile sits naturally between an explicit ordering command and a preference statement.

Citation

@software{echo_dsrn_embed_intent,
  author = {Massimo Roberto Scamarcia},
  title = {Echo-DSRN-v0.1.3-Embed-Intent: Multilingual Intent Classification Embeddings},
  year = {2026},
  url = {https://huggingface.co/Echo-DSRN-v0.1.3-Embed-Intent}
}

---

Note on tokenizer padding

The benchmark results on this card were measured with left padding (padding_side: left), and this model version reproduces them under that convention. A right-padded training version is planned: right padding keeps padded-batch embeddings consistent with single-request embeddings (leading pad tokens do not pollute the recurrent state), so future checkpoints will be batch-composition independent.