SDARZJU-REAL/SDAR
Provides a skill set for on-policy self-distillation (OPSD) and reinforcement learning training of agents,…
No ratings yet
Rate this skill
Stars408
Forks24
Watchers408
Open Issues9
InstallOne-line install
UnconfirmedWhether this repository follows the Agent Skills spec (that is, whether it contains a SKILL.md) has not been verified yet. Until that is settled, clone it and use what you need:
git clone https://github.com/ZJU-REAL/SDAR.git
OverviewOverview
Key Features
Provides a skill set for on-policy self-distillation (OPSD) and reinforcement learning training of agents, packaging steps like policy training, decision alignment, and skill-library management into callable skills. Key capabilities include: 1. AlignOPSD: introduces decision-aligned on-policy distillation beyond timestamps, aligning the distillation target with the agent's actual decisions to reduce policy drift; 2. PR-OPD: performs on-policy distillation with privileged representations, compressing information available only during training into the student policy to improve inference-time performance; 3. UniOPSD: unifies multiple on-policy distillation variants behind one training interface for easy comparison and switching; 4. RetireOPD: judges which skill-library entries to retire or keep, automatically cleaning up inefficient or outdated skills; 5. TASPO / AHEAD / ICSD / BCSD / AgentOPSD / OCSD: cover training paths such as trajectory-aware policy optimization, look-ahead training, in-context skill distillation, behavior-cloning-style distillation, agent-oriented OPSD, and online curriculum distillation.
Use Cases
training and tuning agents that carry a skill library, distilling large-model capabilities into smaller policies, decision alignment within reinforcement learning pipelines, and maintaining skill-library additions, removals, and version iterations.
Target Roles
AI / Algorithm Engineer, Backend Developer, Full-stack / Indie Developer.
TopicsTopics
RatingUser ratings
Nobody has rated this skill yet. Yours would be the first on the site — your vote becomes the initial average.
RepositoryRepository
| Skill ID | 2046 |
|---|---|
| Data source | GitHub |
| Repository | ZJU-REAL/SDAR |
| Category | AI & Agents / Agent Training & Alignment |
| Primary language | Python |
| License | Apache License 2.0 |
| Default branch | master |
| GitHub repository ID | 1238996909 |
| Node ID | R_kgDOSdmXrQ |
| Rank | #1519 (position within the source ranking) |
| Site-wide popularity rank | #1718 (computed live from Stars, across all 5833 entries) |
GrowthStar / Fork growth
One snapshot per day, tracking how Stars and Forks changed
- Now408
- Previous point · 09-24 393
- vs previous point +15
- 2 data points 17 days
- Now24
- Previous point · 09-24 24
- vs previous point ±0
- 2 data points 17 days
TimelineTimeline
Repository created2026-05-14 16:54:036 months ago
Data last updated2026-10-26 12:15:4122 days ago
Last commit2026-10-06 09:23:561 months ago
CommitLatest commit
Version [latest] committed by lll6gg. Update Category: Feature Addition & Enhancement. Commit details: Add recent works powered by SDAR to README. This update introduces new capabilities and module extensions.