AIは変える。知識は引き継ぐ。
Obsidian Agent Kit by UPHASH
トークン削減:Codex 53.2%/Claude Code 48.0%/Cursor 40.7%
※直接読込比、検証3条件中の各ツール最大値。
測定条件・全結果↓
github.com/UpHash-Network/ob…
🤖 Made with AI
Switch AI. Keep your team's knowledge.
Obsidian Agent Kit by UPHASH
Token savings vs direct reads—each tool's best of 3 workloads:
Codex 53.2% · Claude Code 48.0% · Cursor 40.7%
Methods + all results:
github.com/UpHash-Network/ob…
🤖 Made with AI

ALT UPHASH Obsidian Agent Kit: shared Obsidian Markdown and private Git knowledge for Codex, Claude Code and Cursor. Read relevant sections, reuse context and check updates. Best token reductions across three tested workloads vs direct reads: Codex 53.2% (3 projects), Claude Code 48.0% (1 project), Cursor 40.7% (3 projects). Fictional English notes, two four-turn runs per condition, Mac CLI; cached input plus output. Selected maxima; other conditions can increase usage. Full methods are linked in the post.
Heading to #SIGGRAPHAsia2026? 👋
Let's connect in Kuala Lumpur and discover the future of computer graphics together!
📅 1–4 Dec 2026
invt.io/1txbnwqllsl
1/5 LLMの「安定した答え」は、正しい答えとは限りません。
著者の大塩悠貴(UPHASH)です。候補確率と型付き判断を確認できるMini Jevを開発しました。査読前原稿・コード:
uphash-network.github.io/min…
1
36
5/5 検証環境はMac・64GB。モデルは約20.4GBで、軽量動作を保証するものではありません。日本語タスクでの評価です。再現時に詰まる点や、皆さんの具体的な用途を教えてください。
uphash-network.github.io/min…
57
Mini Jev is open source: a local API reading frozen Qwen logits for choices, true/false scores, and ordinal scores without generating answer text. Includes 2,400 self-authored Japanese eval items and optional head-training experiments.
github.com/UpHash-Network/mi…
Developed with AI agents. The 2,400-item suite is self-authored, not an external benchmark. In a separate 1.5B experiment, residual-head training reduced held-out accuracy from 73/96 to 67/96. Matched generation baselines remain future work.
1
23
Mini Jevを公開しました。LLMから文章を生成せず、選択肢・真偽・段階スコアを返すAPIです。自分のデータで補正headを学習・評価するツールも同梱。自作日本語2,400問の結果と失敗例を公開しています。本家Jevとは独立した実装です。
github.com/UpHash-Network/mi…
18



