SHUBHAMOS/meta-pytorch-hackathon
0
1�
2 h��ig � �H � S r SSKJr SSKJrJr SSKJrJr " S S5 r g) u+ 3SHUBHAMOS: AI Email Operations & Triage Environment4reward.py — Dense reward engine (Phase 2)5 6Computes step-level rewards based on agent actions vs. ground truth labels.7Attached to EmailTriageEnv via env.attach_reward_engine(RewardEngine()).8 9Reward design:10 Positive signals (agent did something correct):11 +0.3 — correct category classification12 +0.3 — correct priority assignment13 +0.5 — email resolved successfully14 +0.4 — urgent complaint resolved (bonus)15 +0.1 — relevant reply drafted for billing/urgent16 17 Negative signals (agent did something wrong):18 -0.2 — wrong category19 -0.2 — wrong priority20 -0.5 — urgent complaint ignored21 -0.3 — spam resolved (wasted effort — should be ignored)22 -0.01 — per-step cost (encourages efficiency)23� )� annotations)�List�Optional� )�Action�Emailc � � \ rS rSrSrSrSrSrSrSr Sr24Sr Sr S r
S25rSS jrSS jrSS
jrSS jrSS jrSS jr\SS j5 rSrg)� RewardEngine� zr26Stateless step-level reward calculator.27 28Usage:29 engine = RewardEngine()30 env.attach_reward_engine(engine)31g333333�?g������ɿg �?g�������?g �g333333ӿg�������?g{�G�z��c �� � U R nU R X!R 5 nUc U$ UR S:X a X0R X5 -
nO�UR S:X a X0R X5 -
nOkUR S:X a X0R
U5 -
nOGUR S:X a X0R U5 -
nO#UR S:X a X0R U5 -
n[ US5 $ )z�32Calculate the reward for a single action.33 34Args:35 action: the action just dispatched36 emails: the full email list (with ground truth)37 38Returns:39 float reward value (may be positive, negative, or zero)40�classify_email� set_priority�
mark_resolved� ignore_email� draft_reply� )41� STEP_COST�_find�email_id� action_type�_score_classify�_score_priority�_score_resolve�
_score_ignore� _score_reply�round)�self�action�emails�reward�targets �server/reward.py� score_action�RewardEngine.score_action9 s� � � � � �����F�O�O�4�� �>��M� � � �!1� 1� �*�*�6�:� :�F�
�
�
�>�
1� �*�*�6�:� :�F�
�
�
�?�
2� �)�)�&�1� 1�F�
�
�
�>�
1� �(�(��0� 0�F�
�
�
�=�
0� �'�'��/� /�F�42 �V�Q� � � c �f � UR UR :X a U R $ U R $ �N)�category� gt_category�CORRECT_CATEGORY�WRONG_CATEGORY�r r �emails r" r � RewardEngine._score_classify` s, � � �?�?�e� /� /� /��(�(� (��"�"�"r% c �f � UR UR :X a U R $ U R $ r' )�level� gt_priority�CORRECT_PRIORITY�WRONG_PRIORITYr, s r" r � RewardEngine._score_prioritye s, � � �<�<�5�,�,� ,��(�(� (��"�"�"r% c �� � U R nUR S:X a U R $ UR S:X a X R -
nUR S; a UR ( a X R43 -
nU$ )N�spam�urgent_complaint��
billing_issuer7 )�
RESOLVE_BONUSr) �RESOLVE_SPAM_PENALTY�URGENT_RESOLVE_BONUS�
reply_drafted� REPLY_BONUS)r r- �rs r" r �RewardEngine._score_resolvej st � � � � �� � � �� &��,�,� ,� � � � 2� 2�
�*�*� *�A� � � � E� E�%�J]�J]�
�!�!� !�A��r% c �^ � UR S:X a U R $ UR S:X a gg)Nr7 r6 g�������?� )r) �IGNORE_URGENT_PENALTY�r r- s r" r �RewardEngine._score_ignore{ s1 � � � � � 2� 2��-�-� -� � � �� &��r% c �< � UR S; a U R $ g)Nr8 rB )r) r> rD s r" r �RewardEngine._score_reply� s � � � � � E� E��#�#� #�r% c �> � U H nUR U:X d M Us $ g r' )�id)r r �es r" r �RewardEngine._find� s" � ��A��t�t�x� ��� � r% � N)r r r � List[Email]�return�float)r r r- r rN rO )r- r rN rO )r rM r �strrN zOptional[Email])�__name__�44__module__� __qualname__�__firstlineno__�__doc__r* r+ r2 r3 r: r< rC r; r> r r# r r r r r � staticmethodr �__static_attributes__rL r% r" r45 r46 ss � �� ���N� ���N��M� �� !�� ���K��I�# �N#�47#�48�"�� �� �r% r49 N)50rU �51__future__r �typingr r �modelsr r r52 rL r% r" �<module>r[ s! ���. #� !� !�r� rr%