SHUBHAMOS/meta-pytorch-hackathon

1name: SHUBHAMOS2version: "1.0.0"3description: >4  AI Email Operations & Triage Environment — a production-grade OpenEnv environment5  simulating real-world email inbox management. AI agents classify, prioritize,6  respond to, and resolve emails across three difficulty tiers.7author: Shubhamos Team8homepage: https://huggingface.co/spaces/shubhamos/email-openenv9task_count: 510 11# ── Action Space ───────────────────────────────────────────────────────────────12action_space:13  type: discrete_structured14  description: >15    Structured actions dispatched by action_type. Each action targets a specific16    email by ID and carries optional payload fields.17  actions:18    - name: classify_email19      fields: [email_id, category]20      description: Assign a category to an email (spam/general_inquiry/billing_issue/urgent_complaint)21    - name: set_priority22      fields: [email_id, level]23      description: Set priority level for an email (low/medium/high)24    - name: draft_reply25      fields: [email_id, text]26      description: Draft a reply text to an email (required before mark_resolved on some tasks)27    - name: mark_resolved28      fields: [email_id]29      description: Mark an email as fully resolved30    - name: escalate_email31      fields: [email_id]32      description: Escalate email to human operator33    - name: ignore_email34      fields: [email_id]35      description: Intentionally skip an email (penalized if urgent)36 37# ── Observation Space ──────────────────────────────────────────────────────────38observation_space:39  type: structured40  description: >41    Structured observation returned after each step and reset. Agent sees42    all emails with metadata but NOT ground truth labels.43  fields:44    - name: emails45      type: List[ObservationEmail]46      description: Inbox state — each email has id, subject, sender, sentiment, body_preview (200 chars), agent-assigned category/priority, status flags47    - name: total_emails48      type: int49      description: Total emails in this episode50    - name: pending_count51      type: int52      description: Emails not yet terminally acted on53    - name: resolved_count54      type: int55      description: Emails marked resolved56    - name: escalated_count57      type: int58      description: Emails escalated59    - name: step_count60      type: int61      description: Steps elapsed so far62    - name: max_steps63      type: int64      description: Maximum allowed steps before forced termination65    - name: elapsed_ratio66      type: float67      description: step_count / max_steps (0.0 to 1.0)68 69# ── Reward ─────────────────────────────────────────────────────────────────────70reward_range:71  low: -100.072  high: 100.073reward_description: >74  Dense reward signal. Positive: correct classification (+0.3), correct priority (+0.3),75  resolved email (+0.5), urgent email resolved early (+0.4 bonus). Negative: wrong76  classification (-0.2), wrong priority (-0.2), ignoring urgent (-0.5), step delay (-0.01).77 78# ── Tasks ──────────────────────────────────────────────────────────────────────79tasks:80  - id: peaceful81    name: Peaceful Inbox82    difficulty: peaceful83    grader: server.graders.peaceful_grader:grade84    description: >85      3 emails with clear, unambiguous categories.86      Baseline verification task.87    seed: 188    email_count: 389    max_steps: 2090 91  - id: easy92    name: Basic Inbox93    difficulty: easy94    grader: server.graders.easy_grader:grade95    description: >96      5-8 emails with clear, unambiguous categories. Minimal noise.97      Good starting point for agent calibration.98    seed: 4299    email_count: 8100    max_steps: 50101    expected_categories: [spam, general_inquiry, billing_issue, urgent_complaint]102 103  - id: medium104    name: Support Queue105    difficulty: medium106    grader: server.graders.medium_grader:grade107    description: >108      15-25 emails with mixed priorities and categories. Agent must sequence109      actions effectively and handle urgency.110    seed: 123111    email_count: 20112    max_steps: 120113 114  - id: hard115    name: Enterprise Inbox116    difficulty: hard117    grader: server.graders.hard_grader:grade118    description: >119      40-60 emails with high ambiguity, spam noise, urgent complaints,120      and billing issues. Requires prioritization under pressure.121    seed: 999122    email_count: 50123    max_steps: 300124 125  - id: extreme126    name: Extreme Chaos127    difficulty: extreme128    grader: server.graders.extreme_grader:grade129    description: >130      100 emails under crisis conditions. Total chaos.131      Tests max throughput and prioritization.132    seed: 666133    email_count: 100134    max_steps: 500135 136# ── Endpoints (HTTP API) ───────────────────────────────────────────────────────137endpoints:138  reset: POST /reset139  step: POST /step140  state: GET /state141  health: GET /health142 143# ── Grading Formula ────────────────────────────────────────────────────────────144grading:145  formula: >146    score = classification_accuracy * 0.3147          + priority_accuracy * 0.3148          + resolution_rate * 0.3149          + urgent_handling_score * 0.1150  range: [0.0, 1.0]151  deterministic: true152