ruri-v3-70m

Read the disclaimer below before using this model.

---

ruri-v3-70m -- ONNX for Teradata BYOM

This repository hosts an ONNX-converted version of the upstream model cl-nagoya/ruri-v3-70m, packaged for the Teradata Vantage mldb.ONNXEmbeddings BYOM function. It is not the original PyTorch model -- only the inference graph and tokenizer needed for in-database embedding generation.

What's different from upstream:

• Format: ONNX (opset 14, IR version 8 -- BYOM 6+ compatible), produced from the upstream weights with architecture-aware post-processing baked in.
• Precision: dynamic int8 quantization. See the variants table below for what is shipped for this model.
• Pooling and post-processing: this graph emits the raw sentence_embedding tensor. Pooling rule is mean and the model expects a query-time instruction prefix (see "Instruction prefix" below).
• Verification: every variant's cosine fidelity vs. the upstream PyTorch reference is recorded on a fixed JMTEB retrieval (2 of 8 subsets: jagovfaqs_22k, nlp_journal_title_abs) sample. Numbers may not generalize to your data.

Model details

<details> <summary>Full language list (1)</summary>

• ja

</details>

Instruction prefix

This model was trained with a role-specific prefix scheme: the text you encode must be prefixed according to what you are using the embedding for. The prefix wording is not customisable -- the model only understands these specific strings, and the trailing : (ASCII colon + ASCII space) is part of the prefix.

The two sides of a retrieval pair take different prefixes. This is the part most easily got wrong: prefixing both sides with the same string, or prefixing only one side, degrades retrieval quality without producing any error. Prefix your queries with 検索クエリ: and the passages you index with 検索文書: .

The ONNX graph itself is prefix-agnostic -- the prefix is plain text that flows through the tokenizer. Downstream BYOM SQL is responsible for prepending it, typically with a CTE that concatenates the prefix onto each input row.

Worked example -- the same subject encoded on each side of a retrieval pair:

query:    検索クエリ: 瑠璃色はどんな色？
document: 検索文書: 瑠璃色（るりいろ）は、紫みを帯びた濃い青。

Example SQL (queries and documents prefixed separately, then embedded through the same model):

WITH prefixed_queries AS (
  SELECT id,
         '検索クエリ: ' || query_text AS txt
  FROM   my_query_table
),
prefixed_documents AS (
  SELECT id,
         '検索文書: ' || body_text AS txt
  FROM   my_document_table
)
SELECT *
FROM   mldb.ONNXEmbeddings(
         ON prefixed_queries
         ON embeddings_models     AS ModelTable DIMENSION
         ON embeddings_tokenizers AS TokenizerTable DIMENSION
         USING
           Accumulate('id')
           ModelOutputTensor('sentence_embedding')
       ) AS s;

See cl-nagoya/ruri-v3-70m for the canonical guidance.

Quantization variants

This repository ships the following variants. The Size column is the on-disk size of the ONNX weight file in megabytes (MB, 10^6 bytes). Quality numbers may not generalize to your data; they come from a fixed sample of JMTEB retrieval (2 of 8 subsets: jagovfaqs_22k, nlp_journal_title_abs).

Those numbers were measured with inputs tokenized at a maximum sequence length of 512 tokens.

How to read the quality columns:

• p50 cosine is the median cosine similarity between this variant's embeddings and the fp32 ONNX reference, computed over a fixed evaluation set. Higher means closer to the unquantized model; 1.0 is identical. On the fp32 row the comparison is against the upstream PyTorch model instead, so that row measures export drift rather than quantization drift.
• R@1 is absolute top-1 retrieval accuracy on that evaluation sample: each query is ranked against the whole document pool, and R@1 is the fraction of queries whose own canonical document comes back first. Higher is better. It is not a comparison against fp32 -- every row is measured exactly the same way.
• Δ R@1 vs fp32 is that row's R@1 minus the fp32 row's, on the same corpus at the same sequence length. Because every row is measured the same way, the fp32 row is this model's unquantized ceiling, and this column is the retrieval quality actually given up to quantization. Read it when choosing between variants: two artifacts can post similar absolute R@1 while one has given up several times as much against its own ceiling, and only the delta shows that.

Notes:

• fp32: full-precision reference. Useful for an accuracy ceiling, but BYOM users almost always want one of the int8 variants for in-database scoring -- they are 3-4x smaller and load much faster.
• per_channel: dynamic int8 with weights quantized per output channel. Each output channel keeps its own scale, so layer-wide outliers don't blow up the quantization range. The artifact is roughly 4x smaller than fp32 and is the right default when storage, memory, or load time matters more than the last percent of retrieval quality.
• ffn_skip: dynamic int8 with the feed-forward (FFN) MatMul layers kept in fp32, while attention and projection MatMuls stay quantized. The FFN layers are where most of the quantization error in transformer blocks concentrates; leaving them in fp32 recovers most of the quality loss for a modest size increase. The artifact is roughly 3x smaller than fp32 (larger than the per_channel int8 sibling).

Quickstart: using this model with Teradata BYOM

Requires Teradata Vantage with BYOM 6+ (mldb.ONNXEmbeddings).

import getpass
import teradataml as tdml
from huggingface_hub import hf_hub_download

repo_id   = "Teradata/ruri-v3-70m"
model_id  = "ruri-v3-70m"        # arbitrary, used as the BYOM model_id
onnx_file = "onnx/model-per_channel.onnx"

# 1. Download the ONNX + tokenizer for the chosen variant.
hf_hub_download(repo_id=repo_id, filename=onnx_file,       local_dir="./")
hf_hub_download(repo_id=repo_id, filename="tokenizer.json", local_dir="./")

# 2. Connect to Vantage.
tdml.create_context(
    host=input("host: "),
    username=input("user: "),
    password=getpass.getpass("password: "),
)

# 3. Load model + tokenizer into BYOM tables (one-time per model_id).
tdml.save_byom(model_id=model_id, model_file=onnx_file,
               table_name="embeddings_models")
tdml.save_byom(model_id=model_id, model_file="tokenizer.json",
               table_name="embeddings_tokenizers")

Then call mldb.ONNXEmbeddings against an input table whose txt column carries the strings to embed:

SELECT *
FROM mldb.ONNXEmbeddings(
    ON (SELECT id, txt FROM your_input_table) AS InputTable
    ON (SELECT model_id, model FROM embeddings_models
         WHERE model_id = 'ruri-v3-70m') AS ModelTable DIMENSION
    ON (SELECT model_id, tokenizer FROM embeddings_tokenizers
         WHERE model_id = 'ruri-v3-70m') AS TokenizerTable DIMENSION
    USING
        Accumulate('id')
        ModelOutputTensor('sentence_embedding')
        OutputFormat('FLOAT32(384)')
        OverwriteCachedModel('*')
) AS t
ORDER BY id;

Pooling rule mean is applied inside the converted ONNX graph -- the output tensor named above already contains the pooled, post-processed embedding vector. For instruction-prefix models, prepend the recommended instruction text to each input txt before calling ONNXEmbeddings; the prefix is plain text that the tokenizer handles unchanged.

Original model attribution

The original weights and training methodology belong to the Ruri authors at Nagoya University. Please cite their work, not this repository, in academic contexts. The canonical upstream model card is at cl-nagoya/ruri-v3-70m; refer to it for benchmarks, training details, intended use, and citation information.

cl-nagoya/ruri-v3-70m is itself derived from earlier models, and not every model in that chain is under the same license as this one. The license field in this repository's metadata can only carry a single value (apache-2.0 -- the license of cl-nagoya/ruri-v3-70m), so the full chain is stated here:

Your use of this artifact is subject to the terms of every license in that chain. Check each upstream model card for its authoritative license text.

Reporting issues

For ONNX-conversion or BYOM-compatibility issues specific to this Teradata-converted artifact, please open a Discussion on this model's Hugging Face page. Questions about the underlying model quality, training, or intended use should go to the upstream maintainer's model card.

---

DISCLAIMER: The content herein ("Content") is provided "AS IS" and is not covered by any Teradata Operations, Inc. and its affiliates ("Teradata") agreements. Its listing here does not constitute certification or endorsement by Teradata.

To the extent any of the Content contains or is related to any artificial intelligence ("AI") or other language learning models ("Models") that interoperate with the products and services of Teradata, by accessing, bringing, deploying or using such Models, you acknowledge and agree that you are solely responsible for ensuring compliance with all applicable laws, regulations, and restrictions governing the use, deployment, and distribution of AI technologies. This includes, but is not limited to, AI Diffusion Rules, European Union AI Act, AI-related laws and regulations, privacy laws, export controls, and financial or sector-specific regulations.

While Teradata may provide support, guidance, or assistance in the deployment or implementation of Models to interoperate with Teradata's products and/or services, you remain fully responsible for ensuring that your Models, data, and applications comply with all relevant legal and regulatory obligations. Our assistance does not constitute legal or regulatory approval, and Teradata disclaims any liability arising from non-compliance with applicable laws.

You must determine the suitability of the Models for any purpose. Given the probabilistic nature of machine learning and modeling, the use of the Models may in some situations result in incorrect output that does not accurately reflect the action generated. You should evaluate the accuracy of any output as appropriate for your use case, including by using human review of the output.