dsh-llm-eval
LLM 评估:输出质量评估、幻觉检测、基准测试、回归守护。受 wshobson/agents(38k★ MIT)启发。
88 results
LLM 评估:输出质量评估、幻觉检测、基准测试、回归守护。受 wshobson/agents(38k★ MIT)启发。
Paired experiments and promotion gates for DSH plugins.
Creator-native candidate capture, isolated evaluation, promotion, rollback, and visual lineage for DeepSeek Harness
把技能当可训练参数:像训练神经网络一样训练 agent 技能(epochs/batchsize/learning rate/验证门禁,但不碰模型权重)——rollout→reflect→aggregate→select→update→evaluate 循环、候选编辑仅在严格改善 held-out 验证分时接受、文本学习率预算、零推理时模型调用、部署紧凑 best_skill.md。受 microsoft/SkillOpt(MIT)启发。
DSH plugin: context archiving + historical retrieval — archive_session / search_archive tools over the session-archive skill (zstd session-log decode).
A blind, fair, local Agent arena inside DSH Web: same task, same commit, isolated worktrees, shared verification, judge before you reveal.
Kimi WebBridge host plugin for DeepSeek Harness: model tools that drive the user's REAL browser (navigate, snapshot, click, fill, evaluate, screenshot, network inspection, tab management) through the local kimi-webbridge daemon
DSH-native persistent memory and safe session-aware retrieval plugin.
Evidence packaging, Skill evaluation, and controlled trajectory comparison for DeepSeek Harness sessions
Natural project teaching and first-delivery verification guidance using native DSH instructions.
几何论(共扼谱几何 CSG)知识库插件:纯离线 BM25 检索 208 篇文章全文 + 3833 分块 + 871 条主库真理层,零运行时依赖。Conjugate Spectral Geometry knowledge base for DeepSeek Harness: fully offline BM25 retrieval over 208 articles, 3833 chunks and 871 verified truths.
Cross-session knowledge for DeepSeek Harness: ctx.knowledge service + knowledge_* tools, SQLite FTS5 trigram retrieval (CJK substring) with L1 query expansion. V1.11 contract (write/update/search/list/delete + budget + events + error codes).
DeepSeek Harness plugin: forces one critique-and-improve round after each completed turn.
Independent community DeepSeek Harness plugin for secure multi-file upload, native images, local document retrieval, and coordinate-aware text/PDF/DOCX/XLSX/PPTX reading.
Codex-style lossless-retrieval context manager for the DeepSeek Harness: token-budgeted sliding window, persistent working notes, and cold-history search over the append-only session log
Research lab for DeepSeek Harness: AutoSci-style wiki (experiments/literature/decisions), ASI-Bench-grade eval rigor ledger, arXiv digest, adversarial paper review, and academic writing checks.