Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- models.py +1 -1
- server/rewards.py +3 -1
models.py
CHANGED
|
@@ -109,4 +109,4 @@ class ValidatorState(State):
|
|
| 109 |
correct_reports: int = 0
|
| 110 |
false_positives: int = 0
|
| 111 |
duplicate_reports: int = 0
|
| 112 |
-
score: float = 0.
|
|
|
|
| 109 |
correct_reports: int = 0
|
| 110 |
false_positives: int = 0
|
| 111 |
duplicate_reports: int = 0
|
| 112 |
+
score: float = 0.0001
|
server/rewards.py
CHANGED
|
@@ -91,8 +91,10 @@ def compute_step_reward(
|
|
| 91 |
if is_done_signal:
|
| 92 |
completeness = correct_so_far / max(total_violations, 1)
|
| 93 |
bonus = DONE_BONUS_MULTIPLIER * completeness
|
|
|
|
|
|
|
| 94 |
return RewardBreakdown(
|
| 95 |
-
reward=
|
| 96 |
is_correct=False,
|
| 97 |
is_path_match=False,
|
| 98 |
is_duplicate=False,
|
|
|
|
| 91 |
if is_done_signal:
|
| 92 |
completeness = correct_so_far / max(total_violations, 1)
|
| 93 |
bonus = DONE_BONUS_MULTIPLIER * completeness
|
| 94 |
+
# Clamp to strictly (0, 1) — evaluator requires score never equals 0.0 or 1.0
|
| 95 |
+
bonus = round(max(0.0001, min(0.9999, bonus)), 4)
|
| 96 |
return RewardBreakdown(
|
| 97 |
+
reward=bonus,
|
| 98 |
is_correct=False,
|
| 99 |
is_path_match=False,
|
| 100 |
is_duplicate=False,
|