Model reference · open weights

med-ruri-from-med

Available as managed deployment Embeddings genshiai-daichi · community Embeddings 1 variants 504 dl/mo

med-ruri-from-med is an open-weight embedding model from genshiai-daichi. AxForge deploys and operates it for you on dedicated EU-owned hardware — with the licence handled where one is required.

Available as managed deployment — configured and operated for you on dedicated EU hardware, quoted per deployment.

What it is

Released bygenshiai-daichi
TypeEmbedding models
TaskEmbeddings
Parameters (lead)315M
Context8k tokens
Runs withsentence-transformers
Based ongenshiai-daichi/med-ruri-v3-310m
Released2026-06-14
Popularity504 downloads / month
LicenceOpen weights

About

What med-ruri-from-med is

日本語医療ドメインの文埋め込みモデル (ruri-v3 / ModernBERT-Ja ベース)。学習進行中の自動アップロードです。

このモデルの位置づけ

継続学習: 既に医療 fine-tune 済みの genshiai-daichi/med-ruri-v3-310m(v1) の上に v2 レシピで追加学習

  • 現在の step: 9238 / in-domain nDCG@10: 0.5348
  • 各 step は step-9238revision ブランチで固定取得可。main は最新/最終。

性能 (in-domain nDCG@10)

※ 同一ガイドライン分布の held-out クエリ + 3万 distractor での検索性能汎化/temporal の優劣はこの指標では測れない (別途 ood_affinity で評価)。

Read the full model card
  • baseline (学習前): 0.4244
  • final: 0.5348 (Δ +0.1104)
  • best (全step中): 0.5348
stepnDCG@10
5000.4195
10000.4300
15000.4422
20000.4640
25000.4623
30000.4700
35000.4775
40000.4920
45000.4916
50000.5025
55000.4996
60000.5008
65000.5120
70000.5179
75000.5243
80000.5287
85000.5305
90000.5343
92380.5348

ログ / TensorBoard

  • step毎 nDCG: metrics_progress.json
  • TensorBoard event は各 step-N ブランチの runs/ に同梱 (tensorboard --logdir runs で loss/nDCG を可視化可)。

v2 学習レシピ

  • 精度: fp32 重み + TF32 matmul(ModernBERT は bf16 計算で grad nan 崩壊するため。TF32 で安定かつ高速)
  • NFKC 正規化: PDF 由来の康熙部首などの tokenizer 汚染を除去
  • strip-year: ヘッダの発行年【…(YYYY)…】を除去し、埋め込みを版非依存に
  • temporal hard-negative: 旧版チャンクを hard-neg に(年号への過適合=「学習分布に近い版」嗜好を抑制), oversample ×10
  • loss: MultipleNegativesRankingLoss (scale=50, temp=0.02), lr=1e-5
  • prefix (必須): query=検索クエリ: / document=検索文書: (付けないと精度が大きく落ちる)

使い方

from sentence_transformers import SentenceTransformer
m = SentenceTransformer("genshiai-daichi/med-ruri-v3-310m-v2-from-med", revision="step-9238")  # main で最新も可
q = m.encode(["検索クエリ: 心不全の標準治療は"], normalize_embeddings=True)
d = m.encode(["検索文書: 【…】 …本文…"], normalize_embeddings=True)

注意

  • serving 側も train と同じ前処理 (NFKC + strip-year) が必須(train/serve skew 回避)。
  • 学習中の中間 checkpoint です。最終版・各 step の nDCG@10 比較は metrics_progress.json を参照。

From the published model card. Full card on the HuggingFace links in the sidebar.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys med-ruri-from-med for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (med-ruri-from-med below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/embeddings \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"med-ruri-from-med","input":"text to embed"}'

Create an account — your API key is available in the console. 3M free tokens every 30 days with every new account.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms