Skip to content

README_en.md

Latest commit

 

History

History
60 lines (45 loc) · 8.65 KB

README_en.md

File metadata and controls

60 lines (45 loc) · 8.65 KB
Raw

ScienceEvolve Workflow files and reference implementations

← Back to README · English | 简体中文

This directory presents 21 examples corresponding to a pinned SDK source: three starters and 18 workflows from papers or open-source projects, with JSON, requirements, and reference assets. See the SDK repository for SDK execution. In Web / Server, these are static files for reading, citation, and manual reuse. The application has no dedicated example gallery/API, automatic loading, or example packaging.

Manual use in Web

  1. Keep this directory and its Skill resources on the machine running Server. At the repository root, manually run:

    node example/export-web-workflow.mjs \
  example/aris/workflow.json /tmp/aris.web.json

    On Windows, replace /tmp/aris.web.json with a writable path. The output must be a new file; the script does not overwrite files. Replace aris with your chosen directory. Original workflow.json and requirement.txt remain unchanged.

  2. The exported copy adds the Web orchestration format, removes top-level Git and SDK deployment fields, resolves relative Skill references to actual SKILL.md paths on this machine, and preserves roles, instructions, models, Harnesses, paths, Loops, and inline Skill/MCP definitions. Check original model and reasoning options against your gateway.

  3. On the home page, click “导入 SDK workflow.json” (import SDK workflow.json) and select the exported file. The imported original is read-only, with canvas/JSON views and download. Close the preview, then select it under “快速发起研究任务 → Agent 配置预设” (quick research task → Agent configuration preset) to load an editable copy using Web-host rules.

  4. Copy the corresponding requirement.txt separately into the task requirements and add real task inputs. Choose an installed Harness, available gateway, model, and reasoning options per Agent. Git configuration is not part of the imported preset. If the original has git.enabled: true, configure a Web GitHub account, enable parallel collaboration, and inspect collaboration groups.

  5. Click “提交并继续” (submit and continue) to start. If input uploads are needed, first use “保存 → 另存为新配置” (save → save as new configuration) in quick start, then select that configuration and upload files on the new-project page. Imported originals cannot be overwritten; edited copies can be saved as configurations or task drafts.

For another Server machine, re-export there or update Skill references to actual paths. Copying JSON does not install Skills, MCP, data, or experiment tools. The existing generic runtime resolves selected Skill paths; the application does not register this directory's resources automatically. Remote Web users first download the Server-host export, then select that file in the browser. UI labels above retain the current Chinese interface wording for identification.

Full Web tutorial · Detailed configuration · 21 actual canvases

Workflow examples

Example Type Flow
Default research workflow · JSON · Canvas Starter Ideation, execution, writing, review, and Main coordination
AHE · Code evolution · JSON · Canvas Starter Execution, scoring, refinement Loop, and batch decisions
Software development and testing · JSON · Canvas Starter Coding → Testing → repair or delivery
ARIS · JSON · Canvas Method workflow W1 → W1.5 → W2 → writing and cross-model review
Meta-Harness · JSON · Canvas Method workflow Harness candidates, full-dataset evaluation, and selection
autoresearch · JSON · Canvas Method workflow Training experiments and Git keep/discard
FunSearch · JSON · Canvas Method workflow Multi-island program database, sampling, and actual evaluation
Agent Laboratory · JSON · Canvas Method workflow Seven research substages and solver feedback
SciAgents · JSON · Canvas Method workflow Graph reasoning, hypothesis refinement, and independent review
Robin · JSON · Canvas Method workflow Candidate screening, live-experiment handoff, and analysis
Google AI co-scientist · JSON · Canvas Method workflow Hypothesis generation, reflection, ranking, and evolution
Biomni A1 / E1 · JSON · Canvas Method workflow Resource retrieval and real biological tool feedback
AgentRxiv · JSON · Canvas Method workflow Laboratory research and shared-preprint accumulation
RRSI · JSON · Canvas Method workflow Harness diagnosis, measurement, and regularized selection
DrugEvolve · JSON · Canvas Method workflow Generation, training, debugging, scoring, and archiving
ScientistTwo · JSON · Canvas Method workflow Idea evolution, experiments, writing, and review
AI Scientist v2 · JSON · Canvas Workflow adaptation Ideation, experiment management, research branches, and writing/review nodes
InternAgent 1.5 · JSON · Canvas Workflow adaptation Research planning, generation, verification, and evolution
Biomni-R0 · JSON · Canvas Workflow adaptation Data preparation, trajectories, rewards, training, deployment, and evaluation nodes
Chemical Coscientist · JSON · Canvas Workflow adaptation Planner schedules search, documentation, code, and experiment nodes
Raven Host · JSON · Canvas Workflow adaptation Two specialist slots and batch decisions

Examples are included for their workflow structure and orchestration. Individual directories record mappings and adaptation scope; their original implementation notes and requirements retain their source language.

Files and provenance

Each directory retains workflow.json, requirement.txt, and implementation notes. Named methods include Skills, scripts, reference implementations, licenses, and source metadata. Original graphs and execution protocols target SDK; Web copies follow main scheduling/session rules. See Web / SDK boundaries.

The pinned SDK source is feat/sdk-only-v0.1.1 commit 4160a2cf0259b2a1ed9dc4eb36bc0484d2527a42. catalog.json records directories and status for reference/maintenance; the application does not load it. Project-specific branding uses ScienceEvolve; third-party licenses and source hashes are retained.

Validate original workflows, importable configurations, offline protocols, and scientific outcomes separately. Optional offline checks require Python 3.10+. Install example/tests/requirements.txt in an isolated environment, then run python -m unittest discover -s example/tests. These checks do not invoke research models or GPU experiments.