CoolFace
Apppublic

gsraastha5/Email-openEnv

sourceHugging Facemitupdated 6mo agoView on Hugging Face
0likes
App README

Inbox Triage Action Environment

An OpenEnv-compatible reinforcement learning environment where an AI agent learns to take realistic inbox triage actions such as mark_spam, escalate, or promotions_tab.

Motivation

Email overload is a real-world problem in both personal and workplace inboxes. This environment simulates a more realistic inbox assistant workflow: instead of only classifying messages, the agent must decide what operational action should be taken on each email, such as escalating important messages, routing promotions away from the main inbox, or marking suspicious emails as spam.

Action Space

ActionDescription
mark_spamRoute suspicious, scam, phishing, or malicious emails to spam
escalatePrioritize and surface urgent, important, or operational emails
promotions_tabRoute low-priority commercial and marketing emails to promotions

Observation Space

FieldTypeDescription
email_textstringBody text of the current email
senderstringSender email address
sender_domainstringSender email domain
subjectstringSubject line of the email
contains_linksbooleanWhether the email appears to contain a link or call-to-click
contains_urgency_wordsbooleanWhether the email contains urgent or high-pressure wording
is_external_senderbooleanWhether the sender is outside trusted internal domains
stepintCurrent step number in the episode
total_emailsintTotal emails in the current task
taskstringCurrent task difficulty (easy, medium, hard)
task_descriptionstringHuman-readable description of the task

Tasks

TaskEmailsDifficultyDescription
easy5EasyPersonal inbox triage with obvious spam, promotions, and important messages
medium8MediumWorkplace inbox triage with invoices, deadlines, renewals, and mixed-intent emails
hard10HardSecurity-sensitive inbox triage with phishing-like, ambiguous, and high-stakes business emails

Reward Logic

The environment gives reward at every step, not only at the end.

  • —+2.0 for a correct triage action
  • —stronger negative rewards for unsafe or costly mistakes
  • —-0.1 step penalty to encourage efficient behavior
  • —+10.0 bonus for a perfect episode

Examples of asymmetric penalties:

  • —marking an important email as spam receives a stronger penalty
  • —escalating a suspicious phishing email also receives a strong penalty
  • —misrouting a promotion is penalized less severely than missing an important message

Grading

Each task is graded programmatically using triage accuracy:

text score = correctpredictions / totalpredictions

Setup Instructions

bash:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

## Usage Instructions
Run the API server:

uvicorn env:app --host 0.0.0.0 --port 7860

Open the local docs page:

http://127.0.0.1:7860/docs

Run the baseline inference script:

export APIBASEURL="https://api.openai.com/v1" export MODELNAME="gpt-4o-mini" export HFTOKEN="yourtokenhere" python inference.py

Build and run with Docker:

docker build -t email-openenv .
docker run -p 7860:7860 email-openenv

## Baseline Results

Example output from `python inference.py`:

[START] task=easy env=email-openenv model=gpt-4o-mini [STEP] step=1 action=markspam reward=1.90 done=false error=null [STEP] step=2 action=escalate reward=1.90 done=false error=null [STEP] step=3 action=promotionstab reward=1.90 done=false error=null [STEP] step=4 action=escalate reward=1.90 done=false error=null [STEP] step=5 action=mark_spam reward=11.90 done=true error=null [END] success=true steps=5 score=0.9999 rewards=1.90,1.90,1.90,1.90,11.90

[START] task=medium env=email-openenv model=gpt-4o-mini [STEP] step=1 action=markspam reward=1.90 done=false error=null [STEP] step=2 action=escalate reward=1.90 done=false error=null [STEP] step=3 action=escalate reward=1.90 done=false error=null [STEP] step=4 action=promotionstab reward=1.90 done=false error=null [STEP] step=5 action=promotionstab reward=1.90 done=false error=null [STEP] step=6 action=promotionstab reward=1.90 done=false error=null [STEP] step=7 action=promotions_tab reward=-1.35 done=false error=null [STEP] step=8 action=escalate reward=1.90 done=true error=null [END] success=true steps=8 score=0.8750 rewards=1.90,1.90,1.90,1.90,1.90,1.90,-1.35,1.90

[START] task=hard env=email-openenv model=gpt-4o-mini [STEP] step=1 action=promotionstab reward=1.90 done=false error=null [STEP] step=2 action=promotionstab reward=1.90 done=false error=null [STEP] step=3 action=escalate reward=1.90 done=false error=null [STEP] step=4 action=escalate reward=1.90 done=false error=null [STEP] step=5 action=escalate reward=1.90 done=false error=null [STEP] step=6 action=markspam reward=1.90 done=false error=null [STEP] step=7 action=promotionstab reward=1.90 done=false error=null [STEP] step=8 action=markspam reward=1.90 done=false error=null [STEP] step=9 action=markspam reward=1.90 done=false error=null [STEP] step=10 action=escalate reward=11.90 done=true error=null [END] success=true steps=10 score=0.9999 rewards=1.90,1.90,1.90,1.90,1.90,1.90,1.90,1.90,1.90,11.90```