Spaces:
Running
Running
Commit ·
e44a740
1
Parent(s): 6a0c34c
fix: remove all 0.0/1.0 references, update reward ranges throughout
Browse files
README.md
CHANGED
|
@@ -34,7 +34,7 @@ Every software team triages dozens of bug reports weekly. Getting prioritization
|
|
| 34 |
| `bug_report` | BugReport | Title, body, author, comments |
|
| 35 |
| `task_id` | string | Current difficulty: easy / medium / hard |
|
| 36 |
| `score` | float | Cumulative score this episode |
|
| 37 |
-
| `reward` | float | Reward from last action (0.
|
| 38 |
| `feedback` | string | Human-readable grader feedback |
|
| 39 |
| `done` | bool | Episode complete flag |
|
| 40 |
|
|
@@ -42,25 +42,29 @@ Every software team triages dozens of bug reports weekly. Getting prioritization
|
|
| 42 |
|
| 43 |
### Task 1 — Easy (Priority labeling)
|
| 44 |
Agent assigns a single P0–P3 priority to a bug report.
|
| 45 |
-
- Grader: exact match =
|
| 46 |
- Grader weight: priority 100%
|
|
|
|
| 47 |
|
| 48 |
### Task 2 — Medium (Priority + labels + team)
|
| 49 |
Agent assigns priority, category labels, and team routing.
|
| 50 |
- Grader: priority 45% + label Jaccard similarity 40% + team routing 15%
|
|
|
|
| 51 |
|
| 52 |
### Task 3 — Hard (Full triage)
|
| 53 |
Agent must assign priority, labels, team, and milestone. Security escalation failures are penalized.
|
| 54 |
- Grader: priority 35% + labels 30% + team 20% + milestone 15%
|
| 55 |
- Penalty: −0.15 for missing security escalation
|
|
|
|
| 56 |
|
| 57 |
## Reward function
|
| 58 |
|
| 59 |
Rewards are provided at every step (not just end of episode):
|
| 60 |
-
- Partial credit for close-but-not-exact priority (0.5 vs 0.
|
| 61 |
- Label overlap via Jaccard similarity (continuous signal)
|
| 62 |
- Team routing accuracy (binary, but weighted)
|
| 63 |
- Security escalation penalty discourages ignoring critical signals
|
|
|
|
| 64 |
|
| 65 |
## Setup
|
| 66 |
|
|
@@ -103,13 +107,24 @@ Evaluated with `meta-llama/Llama-3.3-70B-Instruct` via HuggingFace router (tempe
|
|
| 103 |
|
| 104 |
| Task | Score |
|
| 105 |
|------------|-------|
|
| 106 |
-
| Easy |
|
| 107 |
| Medium | 0.500 |
|
| 108 |
-
| Hard |
|
| 109 |
-
| **Avg** | **0.
|
| 110 |
|
| 111 |
Scores vary per run due to random bug sampling from a pool of 5 bugs per task.
|
| 112 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 113 |
## Project structure
|
| 114 |
|
| 115 |
```
|
|
@@ -119,7 +134,7 @@ bug-triage-env/
|
|
| 119 |
│ ├── environment.py # BugTriageEnvironment core logic
|
| 120 |
│ ├── task.py # Bug reports + graders
|
| 121 |
│ └── requirements.txt
|
| 122 |
-
├── model.py #
|
| 123 |
├── client.py # HTTP client
|
| 124 |
├── baseline.py # Groq development script
|
| 125 |
├── inference.py # OpenAI client submission script
|
|
|
|
| 34 |
| `bug_report` | BugReport | Title, body, author, comments |
|
| 35 |
| `task_id` | string | Current difficulty: easy / medium / hard |
|
| 36 |
| `score` | float | Cumulative score this episode |
|
| 37 |
+
| `reward` | float | Reward from last action (0.05–0.95) |
|
| 38 |
| `feedback` | string | Human-readable grader feedback |
|
| 39 |
| `done` | bool | Episode complete flag |
|
| 40 |
|
|
|
|
| 42 |
|
| 43 |
### Task 1 — Easy (Priority labeling)
|
| 44 |
Agent assigns a single P0–P3 priority to a bug report.
|
| 45 |
+
- Grader: exact match = 0.95, one level off = 0.5, else 0.05
|
| 46 |
- Grader weight: priority 100%
|
| 47 |
+
- Reward range: 0.05–0.95
|
| 48 |
|
| 49 |
### Task 2 — Medium (Priority + labels + team)
|
| 50 |
Agent assigns priority, category labels, and team routing.
|
| 51 |
- Grader: priority 45% + label Jaccard similarity 40% + team routing 15%
|
| 52 |
+
- Reward range: 0.05–0.95
|
| 53 |
|
| 54 |
### Task 3 — Hard (Full triage)
|
| 55 |
Agent must assign priority, labels, team, and milestone. Security escalation failures are penalized.
|
| 56 |
- Grader: priority 35% + labels 30% + team 20% + milestone 15%
|
| 57 |
- Penalty: −0.15 for missing security escalation
|
| 58 |
+
- Reward range: 0.05–0.95
|
| 59 |
|
| 60 |
## Reward function
|
| 61 |
|
| 62 |
Rewards are provided at every step (not just end of episode):
|
| 63 |
+
- Partial credit for close-but-not-exact priority (0.5 vs 0.05 vs 0.95)
|
| 64 |
- Label overlap via Jaccard similarity (continuous signal)
|
| 65 |
- Team routing accuracy (binary, but weighted)
|
| 66 |
- Security escalation penalty discourages ignoring critical signals
|
| 67 |
+
- All scores clamped strictly to (0.05, 0.95)
|
| 68 |
|
| 69 |
## Setup
|
| 70 |
|
|
|
|
| 107 |
|
| 108 |
| Task | Score |
|
| 109 |
|------------|-------|
|
| 110 |
+
| Easy | 0.950 |
|
| 111 |
| Medium | 0.500 |
|
| 112 |
+
| Hard | 0.850 |
|
| 113 |
+
| **Avg** | **0.767** |
|
| 114 |
|
| 115 |
Scores vary per run due to random bug sampling from a pool of 5 bugs per task.
|
| 116 |
|
| 117 |
+
## API Endpoints
|
| 118 |
+
|
| 119 |
+
| Method | Endpoint | Description |
|
| 120 |
+
|--------|------------------|------------------------------------|
|
| 121 |
+
| GET | `/` | Health check |
|
| 122 |
+
| POST | `/reset` | Start new episode for a task |
|
| 123 |
+
| POST | `/step` | Submit triage action |
|
| 124 |
+
| GET | `/state` | Get current episode state |
|
| 125 |
+
| GET | `/tasks` | List all tasks with grader info |
|
| 126 |
+
| GET | `/tasks/{id}` | Get specific task metadata |
|
| 127 |
+
|
| 128 |
## Project structure
|
| 129 |
|
| 130 |
```
|
|
|
|
| 134 |
│ ├── environment.py # BugTriageEnvironment core logic
|
| 135 |
│ ├── task.py # Bug reports + graders
|
| 136 |
│ └── requirements.txt
|
| 137 |
+
├── model.py # Pydantic models
|
| 138 |
├── client.py # HTTP client
|
| 139 |
├── baseline.py # Groq development script
|
| 140 |
├── inference.py # OpenAI client submission script
|