File size: 4,712 Bytes
bd7ba9a d183b62 c5c9f44 c9b7f59 9a75cf3 bd7ba9a c5c9f44 c9b7f59 8ddfad6 bd7ba9a 8ddfad6 c5c9f44 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 c5c9f44 8ddfad6 bd7ba9a 9a75cf3 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a c5c9f44 8ddfad6 c5c9f44 bd7ba9a c5c9f44 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a 8ddfad6 bd7ba9a d183b62 bd7ba9a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 | ---
title: Document Classification Env
emoji: π
colorFrom: blue
colorTo: green
sdk: docker
pinned: false
tags:
- openenv
- document-classification
- reinforcement-learning
- gymnasium
---
# π Document Classification Environment
**OpenEnv submission for Meta x PyTorch Hackathon**
An interactive Gymnasium-compatible environment where AI agents learn to classify and route customer support tickets to the correct department.
---
## π― Real-World Task
Customer support teams receive hundreds of tickets daily β billing issues, technical bugs, HR complaints, legal queries. This environment simulates that routing challenge, training agents to read a document and instantly decide which department should handle it.
**Why this matters:** Misrouted tickets waste time and frustrate customers. A well-trained agent can reduce misrouting by 80%+.
---
## ποΈ Environment Design
```
DocumentClassificationEnv(task_difficulty="hard", seed=42)
βββ observation_space: Dict
β βββ content: Text (the document)
β βββ document_id: Text
β βββ word_count: Box(1,)
β βββ has_urgency_markers: MultiBinary(1)
β βββ features: Box(100,) β TF-IDF features
β βββ document_index: Box(1,)
β βββ total_documents: Box(1,)
βββ action_space: Discrete(N) β N = num categories
```
### API
```python
from environment import DocumentClassificationEnv
env = DocumentClassificationEnv(task_difficulty="hard", seed=42)
obs, info = env.reset()
while True:
action = your_agent(obs) # int: category index
obs, reward, done, _, info = env.step(action)
print(f"Reward: {reward:.3f}, Correct: {info['is_correct']}")
if done:
print(info["episode_summary"])
break
```
---
## π Three Progressive Tasks
| Task | Categories | Documents | Time Limit | Target Score |
|------|-----------|-----------|------------|--------------|
| **Easy** | 5 | 100 | None | 0.85 |
| **Medium** | 10 | 500 | 2 sec | 0.75 |
| **Hard** | 22 | 1000 | 1 sec | 0.75 |
### Categories (Hard Mode β 22 total)
`General` `Billing` `Billing-Dispute` `Billing-Refund` `Support` `Support-Urgent` `Support-Normal` `Technical` `Technical-Bug` `Technical-Feature` `HR` `HR-Payroll` `HR-Benefits` `HR-Complaint` `Legal` `Legal-Contract` `Legal-Compliance` `Executive` `Executive-Strategic` `Finance` `Marketing` `Operations`
---
## π Reward Function
```python
reward = accuracy_reward + speed_bonus
# accuracy_reward:
# +1.0 for correct classification
# -0.5 for wrong classification
# speed_bonus (difficulty-dependent):
# Easy: +0.10 if < 100ms
# Medium: +0.15 if < 200ms, +0.10 if < 500ms
# Hard: +0.20 if < 100ms, +0.10 if < 300ms
```
Partial credit via speed bonus encourages efficient inference, not just accuracy.
---
## π Baseline Results (Keyword Agent)
| Task | Score | Notes |
|------|-------|-------|
| EASY | **0.77** | Simple keyword matching |
| MEDIUM | **0.89** | Priority-ordered keywords |
| HARD | **0.165** | 22 categories, harder to distinguish |
Better agents (TF-IDF similarity, fine-tuned LLM) can significantly beat baseline.
---
## π Quick Start
```bash
git clone https://huggingface.co/spaces/TanujInsane/document-classification-env
cd document-classification-env
pip install -r requirements.txt
# Run baseline
python baseline_inference.py --task all --output results.json
# Launch UI
python app.py
```
---
## π File Structure
```
βββ environment.py # Main Gymnasium environment
βββ tasks.py # Document generation + TF-IDF features
βββ grading.py # Scoring logic (0.0 - 1.0)
βββ baseline_inference.py # Keyword-based baseline agent
βββ app.py # Gradio interactive demo
βββ test_environment.py # 5 unit tests (all passing β
)
βββ openenv.yaml # OpenEnv specification
βββ Dockerfile # Container deployment
βββ requirements.txt # Dependencies
```
---
## π¬ Reproducibility
- Seed-controlled document generation
- Fixed test sets for fair grading
- Deterministic reward calculation
- Docker containerization for consistent deployment
---
## π‘ Improving Beyond Baseline
```python
# Example: TF-IDF similarity agent (beats keyword matching)
from sklearn.metrics.pairwise import cosine_similarity
class TFIDFAgent:
def __init__(self, difficulty):
self.env = DocumentClassificationEnv(difficulty)
# Pre-compute category centroid vectors
# Use cosine similarity at inference time
...
```
---
*Built for Meta x PyTorch OpenEnv Hackathon | Gymnasium-compatible | Docker deployed*
|