Spaces:
Sleeping
Sleeping
Merge branch 'main' of https://huggingface.co/spaces/SheSync/meta-hackathon-project
Browse files- my_env/README.md +68 -21
my_env/README.md
CHANGED
|
@@ -19,6 +19,19 @@ This environment teaches an AI to make that decision smartly by looking at:
|
|
| 19 |
The AI learns by getting **reward scores** β good decisions get high scores, bad decisions get low scores!
|
| 20 |
|
| 21 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
## π± Real World Problem It Solves
|
| 24 |
|
|
@@ -38,16 +51,27 @@ Every day the AI:
|
|
| 38 |
4. **Learns** from the score and does better next time!
|
| 39 |
|
| 40 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
## π§ Actions (
|
| 43 |
|
| 44 |
-
|
| 45 |
-
|--------|------------|-------------|
|
| 46 |
-
| `none` | 0mm | Soil is wet enough or it's raining |
|
| 47 |
-
| `low` | 10mm | Soil is slightly dry |
|
| 48 |
-
| `medium` | 25mm | Soil is moderately dry |
|
| 49 |
-
| `high` | 50mm | Soil is very dry |
|
| 50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
---
|
| 52 |
|
| 53 |
## π Observations (What AI Can See)
|
|
@@ -86,15 +110,17 @@ Every day the AI:
|
|
| 86 |
|
| 87 |
---
|
| 88 |
|
| 89 |
-
## π Scoring
|
| 90 |
|
| 91 |
| Score | Meaning |
|
| 92 |
|-------|---------|
|
| 93 |
-
| 1.0 | Perfect
|
| 94 |
-
| 0.
|
| 95 |
-
| 0.
|
| 96 |
-
|
|
| 97 |
-
|
|
|
|
|
|
|
|
| 98 |
|
| 99 |
**Bonus:** AI gets penalized if it waters crops on rainy days (water waste!)
|
| 100 |
|
|
@@ -120,6 +146,16 @@ http://localhost:8000/docs
|
|
| 120 |
```
|
| 121 |
|
| 122 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 123 |
|
| 124 |
## π§ͺ Test the Endpoints
|
| 125 |
|
|
@@ -140,7 +176,7 @@ curl -X POST http://localhost:8000/step \
|
|
| 140 |
curl http://localhost:8000/tasks
|
| 141 |
```
|
| 142 |
|
| 143 |
-
### Run
|
| 144 |
```bash
|
| 145 |
curl http://localhost:8000/baseline?city=Indore
|
| 146 |
```
|
|
@@ -163,15 +199,24 @@ python baseline.py --city Mumbai
|
|
| 163 |
python baseline.py --city Delhi
|
| 164 |
```
|
| 165 |
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
Task 1 (Basic Moisture Control):
|
| 169 |
-
Task 2 (Weather Adaptive): 0.
|
| 170 |
-
Task 3 (Full Season Management): 0.
|
| 171 |
-
|
| 172 |
-
|
| 173 |
|
| 174 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 175 |
|
| 176 |
## π Project Structure
|
| 177 |
```
|
|
@@ -205,6 +250,8 @@ my_env/
|
|
| 205 |
| `/grader` | POST | Get score for completed episode |
|
| 206 |
| `/baseline` | GET | Run baseline agent, get scores |
|
| 207 |
| `/docs` | GET | Interactive API documentation |
|
|
|
|
|
|
|
| 208 |
|
| 209 |
---
|
| 210 |
|
|
|
|
| 19 |
The AI learns by getting **reward scores** β good decisions get high scores, bad decisions get low scores!
|
| 20 |
|
| 21 |
---
|
| 22 |
+
## π€ LLM Integration
|
| 23 |
+
|
| 24 |
+
This environment uses a Large Language Model (LLM) as the decision-making agent.
|
| 25 |
+
|
| 26 |
+
At each step:
|
| 27 |
+
- The current state is passed to the LLM
|
| 28 |
+
- The LLM reasons about irrigation needs
|
| 29 |
+
- It outputs a structured JSON action
|
| 30 |
+
|
| 31 |
+
This ensures:
|
| 32 |
+
- No hardcoded decision logic
|
| 33 |
+
- Fully AI-driven behavior
|
| 34 |
+
- Generalizable across tasks
|
| 35 |
|
| 36 |
## π± Real World Problem It Solves
|
| 37 |
|
|
|
|
| 51 |
4. **Learns** from the score and does better next time!
|
| 52 |
|
| 53 |
---
|
| 54 |
+
## π RL Interaction Loop
|
| 55 |
+
|
| 56 |
+
1. `reset()` β initialize environment
|
| 57 |
+
2. `state()` β get observation
|
| 58 |
+
3. LLM generates action
|
| 59 |
+
4. `step(action)` β environment updates
|
| 60 |
+
5. reward returned (0β1)
|
| 61 |
+
6. loop continues
|
| 62 |
+
|
| 63 |
+
This follows OpenEnv RL standards.
|
| 64 |
|
| 65 |
+
## π§ Actions (LLM-Driven)
|
| 66 |
|
| 67 |
+
The AI agent generates actions dynamically using an LLM.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
|
| 69 |
+
Each step returns:
|
| 70 |
+
|
| 71 |
+
- `irrigate`: true/false
|
| 72 |
+
- `water_amount`: value between 0β100
|
| 73 |
+
|
| 74 |
+
This enables flexible and intelligent irrigation decisions instead of fixed rules.
|
| 75 |
---
|
| 76 |
|
| 77 |
## π Observations (What AI Can See)
|
|
|
|
| 110 |
|
| 111 |
---
|
| 112 |
|
| 113 |
+
## π Scoring (Normalized Rewards)
|
| 114 |
|
| 115 |
| Score | Meaning |
|
| 116 |
|-------|---------|
|
| 117 |
+
| 1.0 | Perfect |
|
| 118 |
+
| 0.7 | Good |
|
| 119 |
+
| 0.4 | Acceptable |
|
| 120 |
+
| 0.1 | Poor |
|
| 121 |
+
| 0.0 | Failure |
|
| 122 |
+
|
| 123 |
+
All rewards are clamped between 0.0 and 1.0.
|
| 124 |
|
| 125 |
**Bonus:** AI gets penalized if it waters crops on rainy days (water waste!)
|
| 126 |
|
|
|
|
| 146 |
```
|
| 147 |
|
| 148 |
---
|
| 149 |
+
## π€ Run LLM Inference Script
|
| 150 |
+
|
| 151 |
+
Set environment variables first:
|
| 152 |
+
|
| 153 |
+
```bash
|
| 154 |
+
export API_BASE_URL="https://router.huggingface.co/v1"
|
| 155 |
+
export MODEL_NAME="meta-llama/Llama-3.3-70B-Instruct"
|
| 156 |
+
export HF_TOKEN="your_token_here"
|
| 157 |
+
|
| 158 |
+
|
| 159 |
|
| 160 |
## π§ͺ Test the Endpoints
|
| 161 |
|
|
|
|
| 176 |
curl http://localhost:8000/tasks
|
| 177 |
```
|
| 178 |
|
| 179 |
+
### Run LLM evaluation script:
|
| 180 |
```bash
|
| 181 |
curl http://localhost:8000/baseline?city=Indore
|
| 182 |
```
|
|
|
|
| 199 |
python baseline.py --city Delhi
|
| 200 |
```
|
| 201 |
|
| 202 |
+
## π Baseline Results (LLM Agent)
|
| 203 |
+
|
| 204 |
+
Task 1 (Basic Moisture Control): 1.00
|
| 205 |
+
Task 2 (Weather Adaptive): 0.98
|
| 206 |
+
Task 3 (Full Season Management): 0.80
|
| 207 |
+
|
| 208 |
+
Average Score: 0.93
|
| 209 |
|
| 210 |
---
|
| 211 |
+
## βοΈ Environment Variables
|
| 212 |
+
|
| 213 |
+
The following variables are required:
|
| 214 |
+
|
| 215 |
+
- API_BASE_URL β LLM endpoint
|
| 216 |
+
- MODEL_NAME β model used for inference
|
| 217 |
+
- HF_TOKEN β authentication token
|
| 218 |
+
|
| 219 |
+
Set them before running inference or deployment.
|
| 220 |
|
| 221 |
## π Project Structure
|
| 222 |
```
|
|
|
|
| 250 |
| `/grader` | POST | Get score for completed episode |
|
| 251 |
| `/baseline` | GET | Run baseline agent, get scores |
|
| 252 |
| `/docs` | GET | Interactive API documentation |
|
| 253 |
+
| `/ui` | GET | Interactive farm UI |
|
| 254 |
+
| `/inference-step` | POST | Single LLM decision |
|
| 255 |
|
| 256 |
---
|
| 257 |
|