Spaces:
Sleeping
Sleeping
fix: complete HF Space README with full metadata and documentation
Browse files
README.md
CHANGED
|
@@ -4,8 +4,8 @@ emoji: π§
|
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: green
|
| 6 |
sdk: docker
|
| 7 |
-
pinned: false
|
| 8 |
app_port: 7860
|
|
|
|
| 9 |
tags:
|
| 10 |
- openenv
|
| 11 |
- email-triage
|
|
@@ -14,7 +14,193 @@ tags:
|
|
| 14 |
- nlp
|
| 15 |
---
|
| 16 |
|
| 17 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
## Structure
|
| 20 |
|
|
|
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: green
|
| 6 |
sdk: docker
|
|
|
|
| 7 |
app_port: 7860
|
| 8 |
+
pinned: false
|
| 9 |
tags:
|
| 10 |
- openenv
|
| 11 |
- email-triage
|
|
|
|
| 14 |
- nlp
|
| 15 |
---
|
| 16 |
|
| 17 |
+
# π§ Email Triage Environment
|
| 18 |
+
|
| 19 |
+
> **Meta Γ PyTorch OpenEnv Hackathon β Round 1**
|
| 20 |
+
> An RL environment where an AI agent learns to triage real-world emails.
|
| 21 |
+
|
| 22 |
+
## π― What Is This?
|
| 23 |
+
|
| 24 |
+
This environment simulates an **email triage system** β a task every business performs daily. An agent receives a queue of emails and must:
|
| 25 |
+
|
| 26 |
+
1. **Classify** the email (spam / billing / technical / hr / general)
|
| 27 |
+
2. **Assign a priority** (low / medium / high / urgent)
|
| 28 |
+
3. **Route** to the correct department (support / finance / engineering / hr / management)
|
| 29 |
+
4. **Draft a professional reply** (required for medium and hard tasks)
|
| 30 |
+
|
| 31 |
+
Rewards are **continuous and partial** β agents earn credit for getting each dimension right, incentivising learning even when the full answer isn't perfect.
|
| 32 |
+
|
| 33 |
+
---
|
| 34 |
+
|
| 35 |
+
## ποΈ Action Space
|
| 36 |
+
|
| 37 |
+
| Field | Type | Values |
|
| 38 |
+
|--------------|--------|-----------------------------------------------------|
|
| 39 |
+
| `category` | string | `spam` Β· `billing` Β· `technical` Β· `hr` Β· `general` |
|
| 40 |
+
| `priority` | string | `low` Β· `medium` Β· `high` Β· `urgent` |
|
| 41 |
+
| `department` | string | `support` Β· `finance` Β· `engineering` Β· `hr` Β· `management` |
|
| 42 |
+
| `reply` | string | Free-text reply draft (empty string for easy tasks) |
|
| 43 |
+
|
| 44 |
+
---
|
| 45 |
+
|
| 46 |
+
## ποΈ Observation Space
|
| 47 |
+
|
| 48 |
+
| Field | Type | Description |
|
| 49 |
+
|-------------------|--------|----------------------------------------------|
|
| 50 |
+
| `email_id` | string | Unique email identifier |
|
| 51 |
+
| `subject` | string | Email subject line |
|
| 52 |
+
| `body` | string | Full email body |
|
| 53 |
+
| `sender` | string | Sender's email address |
|
| 54 |
+
| `task_level` | string | `easy` Β· `medium` Β· `hard` |
|
| 55 |
+
| `feedback` | string | Natural-language grader feedback |
|
| 56 |
+
| `score` | float | Cumulative average score (0.0β1.0) |
|
| 57 |
+
| `emails_remaining`| int | Number of emails left in the episode |
|
| 58 |
+
| `done` | bool | Episode complete flag |
|
| 59 |
+
| `reward` | float | Per-step reward (0.0β1.0) |
|
| 60 |
+
|
| 61 |
+
---
|
| 62 |
+
|
| 63 |
+
## π Task Descriptions
|
| 64 |
+
|
| 65 |
+
### π’ Easy β Classification Only
|
| 66 |
+
- **5 emails** per episode (spam, billing, technical, HR, general)
|
| 67 |
+
- Agent must correctly classify category, priority, and department
|
| 68 |
+
- No reply required
|
| 69 |
+
- Score: `category(0.4) + priority(0.3) + department(0.3)`
|
| 70 |
+
|
| 71 |
+
### π‘ Medium β Classification + Short Reply
|
| 72 |
+
- **3 emails** per episode (billing urgency, technical incident, HR complaint)
|
| 73 |
+
- Agent must classify AND write a professional 2-4 sentence reply
|
| 74 |
+
- Score: `classification(0.5) + reply_quality(0.5)`
|
| 75 |
+
- Reply graded on: professional tone, keyword coverage
|
| 76 |
+
|
| 77 |
+
### π΄ Hard β Classification + Detailed Accurate Reply
|
| 78 |
+
- **3 emails** per episode (enterprise escalations, security incidents, board complaints)
|
| 79 |
+
- Agent must classify AND write a detailed, accurate reply covering specific required points
|
| 80 |
+
- Score: `classification(0.4) + reply_depth(0.6)`
|
| 81 |
+
- Reply graded on: length, tone, keyword coverage (9 required keywords per email)
|
| 82 |
+
|
| 83 |
+
---
|
| 84 |
+
|
| 85 |
+
## π Reward Function
|
| 86 |
+
|
| 87 |
+
Rewards are **partial and continuous** β agents earn credit at every step:
|
| 88 |
+
|
| 89 |
+
| Condition | Reward |
|
| 90 |
+
|-------------------------------------------|---------------|
|
| 91 |
+
| Correct category | +0.15 β +0.40 |
|
| 92 |
+
| Correct priority | +0.13 β +0.30 |
|
| 93 |
+
| Correct department | +0.12 β +0.30 |
|
| 94 |
+
| Reply provided (non-trivial) | +0.05 β +0.10 |
|
| 95 |
+
| Professional tone in reply | +0.05 β +0.10 |
|
| 96 |
+
| Each required keyword matched | proportional |
|
| 97 |
+
| Missing reply on medium/hard | β0.10 penalty |
|
| 98 |
+
|
| 99 |
+
---
|
| 100 |
+
|
| 101 |
+
## π Quick Start
|
| 102 |
+
|
| 103 |
+
### 1. Install dependencies
|
| 104 |
+
```bash
|
| 105 |
+
git clone https://huggingface.co/spaces/Shivacode/rl-hackathon
|
| 106 |
+
cd rl-hackathon
|
| 107 |
+
pip install -e email_triage_rl_hackathon/
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
### 2. Run the server locally
|
| 111 |
+
```bash
|
| 112 |
+
# Easy task (default)
|
| 113 |
+
uvicorn email_triage_rl_hackathon.server.app:app --host 0.0.0.0 --port 8000 --reload
|
| 114 |
+
|
| 115 |
+
# Medium task
|
| 116 |
+
TASK_LEVEL=medium uvicorn email_triage_rl_hackathon.server.app:app --port 8000
|
| 117 |
+
|
| 118 |
+
# Hard task
|
| 119 |
+
TASK_LEVEL=hard uvicorn email_triage_rl_hackathon.server.app:app --port 8000
|
| 120 |
+
```
|
| 121 |
+
|
| 122 |
+
### 3. Interact with the environment (Python)
|
| 123 |
+
```python
|
| 124 |
+
import asyncio
|
| 125 |
+
from email_triage_rl_hackathon import EmailTriageEnv, TriageAction
|
| 126 |
+
|
| 127 |
+
async def main():
|
| 128 |
+
env = EmailTriageEnv(base_url="http://localhost:8000")
|
| 129 |
+
obs = await env.reset(seed=42, task_level="easy")
|
| 130 |
+
print(obs.subject)
|
| 131 |
+
|
| 132 |
+
result = await env.step(TriageAction(
|
| 133 |
+
category="billing", priority="urgent",
|
| 134 |
+
department="finance",
|
| 135 |
+
reply="Dear customer, we apologize and will investigate within 24 hours."
|
| 136 |
+
))
|
| 137 |
+
print(result.reward, result.observation.feedback)
|
| 138 |
+
await env.close()
|
| 139 |
+
|
| 140 |
+
asyncio.run(main())
|
| 141 |
+
```
|
| 142 |
+
|
| 143 |
+
### 4. Run with Docker
|
| 144 |
+
```bash
|
| 145 |
+
docker build -t email-triage-env .
|
| 146 |
+
docker run -p 7860:7860 -e TASK_LEVEL=easy email-triage-env
|
| 147 |
+
```
|
| 148 |
+
|
| 149 |
+
### 5. Run inference against this Space
|
| 150 |
+
```bash
|
| 151 |
+
HF_TOKEN=hf_your_token_here \
|
| 152 |
+
ENV_BASE_URL=https://Shivacode-rl-hackathon.hf.space \
|
| 153 |
+
TASK_LEVEL=easy \
|
| 154 |
+
python inference.py
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
---
|
| 158 |
+
|
| 159 |
+
## π Scores (Qwen/Qwen2.5-72B-Instruct via HF Router)
|
| 160 |
+
|
| 161 |
+
| Task Level | Steps | Score |
|
| 162 |
+
|------------|-------|--------|
|
| 163 |
+
| Easy | 5 | 0.880 |
|
| 164 |
+
| Medium | 3 | 0.830 |
|
| 165 |
+
| Hard | 3 | 0.800 |
|
| 166 |
+
| **Overall**| | **0.837** |
|
| 167 |
+
|
| 168 |
+
---
|
| 169 |
+
|
| 170 |
+
## ποΈ Project Structure
|
| 171 |
+
|
| 172 |
+
```
|
| 173 |
+
rl-hackathon/
|
| 174 |
+
βββ Dockerfile # HF Spaces Docker deployment (port 7860)
|
| 175 |
+
βββ inference.py # Run agent against this environment
|
| 176 |
+
βββ baseline.py # Baseline script (OpenAI models)
|
| 177 |
+
βββ email_triage_rl_hackathon/
|
| 178 |
+
βββ __init__.py # Public exports
|
| 179 |
+
βββ models.py # TriageAction, TriageObservation
|
| 180 |
+
βββ client.py # EmailTriageEnv (EnvClient)
|
| 181 |
+
βββ openenv.yaml # Environment manifest
|
| 182 |
+
βββ pyproject.toml # Package config + scripts entry point
|
| 183 |
+
βββ uv.lock # Locked dependencies
|
| 184 |
+
βββ server/
|
| 185 |
+
βββ app.py # FastAPI app (uvicorn entry point)
|
| 186 |
+
βββ my_environment.py # EmailTriageEnvironment (reset/step/state)
|
| 187 |
+
βββ email_data.py # 11 curated emails across 3 levels
|
| 188 |
+
βββ graders.py # Deterministic graders (easy/medium/hard)
|
| 189 |
+
βββ requirements.txt
|
| 190 |
+
```
|
| 191 |
+
|
| 192 |
+
---
|
| 193 |
+
|
| 194 |
+
## π Environment Variables
|
| 195 |
+
|
| 196 |
+
| Variable | Default | Description |
|
| 197 |
+
|-----------------|----------------------------|--------------------------------------|
|
| 198 |
+
| `TASK_LEVEL` | `easy` | Task difficulty: easy / medium / hard |
|
| 199 |
+
| `PORT` | `7860` | Server port |
|
| 200 |
+
| `HF_TOKEN` | β | Required for inference.py (HF router)|
|
| 201 |
+
| `MODEL_NAME` | `Qwen/Qwen2.5-72B-Instruct`| Model for inference |
|
| 202 |
+
| `ENV_BASE_URL` | `http://localhost:8000` | Environment server URL |
|
| 203 |
+
|
| 204 |
|
| 205 |
## Structure
|
| 206 |
|