Spaces:
Running
Running
File size: 4,057 Bytes
80d8c84 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 | ---
title: ReplicaLab Architecture
---
flowchart TB
subgraph SCENARIO["Scenario Generation"]
direction LR
PT["Paper Templates<br/><i>Cell Bio · ML Benchmark · Psych</i>"]
CG["Constraint Generator<br/><i>Equipment · Budget · Staff · Calendar</i>"]
SE["Scenario Engine<br/><i>Seed → Deterministic World</i>"]
PT --> SE
CG --> SE
end
subgraph ENV["ReplicaLab Environment (OpenEnv)"]
direction TB
STATE["Environment State<br/><i>Paper · Constraints · Round · Budget<br/>Protocol · History · Done Flag</i>"]
subgraph AGENTS["Agent Loop"]
direction LR
SCI["🔬 Scientist Agent<br/><i>Trainable LLM Policy</i><br/><b>Actions:</b> propose · revise<br/>ask · accept"]
LM["🏗️ Lab Manager Agent<br/><i>Rule-Based Policy</i><br/><b>Actions:</b> report · suggest<br/>reject · accept"]
SCI -- "Proposal /<br/>Question" --> LM
LM -- "Constraint /<br/>Substitution" --> SCI
end
subgraph JUDGE["Judge Engine"]
direction LR
RUBRIC["Rubric Scorer<br/><i>Deterministic</i>"]
EXPLAIN["Explanation Layer<br/><i>Optional LLM</i>"]
RUBRIC --> EXPLAIN
end
STATE --> AGENTS
AGENTS -- "step()" --> STATE
STATE -- "Episode End" --> JUDGE
end
subgraph REWARD["Reward Computation"]
direction LR
R["Rigor<br/>Score"]
FE["Feasibility<br/>Score"]
FI["Fidelity<br/>Score"]
BONUS["Efficiency +<br/>Communication<br/>Bonus"]
PEN["Penalties<br/><i>Timeout · Over Budget<br/>Missing Controls</i>"]
TOTAL["<b>Total Reward</b><br/><i>10 × R × Fe × Fi<br/>+ Bonus − Penalties</i>"]
R --> TOTAL
FE --> TOTAL
FI --> TOTAL
BONUS --> TOTAL
PEN --> TOTAL
end
subgraph TRAINING["RL Training Pipeline"]
direction LR
COLAB["Google Colab<br/><i>TRL / Unsloth · GRPO</i>"]
ROLLOUT["Rollout Loop<br/><i>reset() → step() → reward</i>"]
CURVES["Reward Curves<br/><i>Before vs After</i>"]
COLAB --> ROLLOUT --> CURVES
end
subgraph SERVING["Deployment & Serving"]
direction LR
FASTAPI["FastAPI +<br/>WebSocket Server"]
DOCKER["Docker Container"]
HF["Hugging Face Space<br/><i>sdk: docker · port: 7860</i>"]
FASTAPI --> DOCKER --> HF
end
subgraph UI["Frontend"]
direction LR
REACT["React + Vite UI"]
FALLBACK["OpenEnv /web<br/><i>Fallback</i>"]
subgraph PANELS["Layout"]
direction TB
LEFT["Left Panel<br/><i>Paper · Seed · Round</i>"]
MID["Middle Panel<br/><i>Negotiation Log</i>"]
RIGHT["Right Panel<br/><i>Protocol · Budget · Scores</i>"]
end
REACT --> PANELS
FALLBACK --> PANELS
end
SE -- "reset(seed)" --> ENV
JUDGE -- "Scores" --> REWARD
TOTAL -- "Reward Signal" --> TRAINING
ROLLOUT -- "Episodes" --> ENV
ENV -- "API" --> FASTAPI
FASTAPI -- "WebSocket" --> REACT
FASTAPI -- "WebSocket" --> FALLBACK
TRAINING -. "Updated Scientist<br/>Policy Weights" .-> SCI
classDef scenario fill:#3b82f6,stroke:#1d4ed8,color:#fff
classDef env fill:#1e293b,stroke:#475569,color:#e2e8f0
classDef agent fill:#8b5cf6,stroke:#6d28d9,color:#fff
classDef judge fill:#f59e0b,stroke:#d97706,color:#1e293b
classDef reward fill:#10b981,stroke:#059669,color:#fff
classDef training fill:#ef4444,stroke:#dc2626,color:#fff
classDef serving fill:#6366f1,stroke:#4f46e5,color:#fff
classDef ui fill:#ec4899,stroke:#db2777,color:#fff
classDef panel fill:#fdf2f8,stroke:#ec4899,color:#1e293b
class PT,CG,SE scenario
class STATE env
class SCI,LM agent
class RUBRIC,EXPLAIN judge
class R,FE,FI,BONUS,PEN,TOTAL reward
class COLAB,ROLLOUT,CURVES training
class FASTAPI,DOCKER,HF serving
class REACT,FALLBACK ui
class LEFT,MID,RIGHT panel
|