Hugging Face's logo

Spaces:
YashashMathur
/
aegis_training
Runtime error

App Files Community
aegis_training
File size: 1,324 Bytes
609a677
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
name: AEGIS-ENV
version: 0.1.0
description: "AI Fleet Oversight RL Training Environment - Train an LLM to detect policy violations in AI agent workflows"
url: "https://github.com/mathuryashash/AEGIS-ENV"
tags:
  - reinforcement-learning
  - safety
  - llm-training
  - openenv
  - meta-hackathon
author: "Hitanshu & Yashash"
interface:
  type: fastapi
  port: 7860
metrics:
  track_suspicion: true
  enable_episode_logs: true
  log_rewards: true
environment:
  type: rl
  observation:
    - worker_cot_trace
    - worker_output
    - policies
    - memory_context
    - simulation_logs
  actions:
    - decision: ALLOW|BLOCK|ESCALATE
    - violation_type: string
    - policy_rule_cited: string
    - explanation: string
  rewards:
    - decision_correctness
    - violation_identification
    - policy_citation
    - explanation_quality
    - format_compliance
rubric:
  decision_accuracy:
    weight: 0.35
    description: "Correct ALLOW/BLOCK/ESCALATE classification"
  violation_detection:
    weight: 0.25
    description: "Accurate violation type identification"
  policy_reasoning:
    weight: 0.20
    description: "Citing correct policy rules"
  explanation_quality:
    weight: 0.10
    description: "Clear reasoning in explanation"
  format_compliance:
    weight: 0.10
    description: "Valid JSON output format"