diff options
Diffstat (limited to 'collaborativeagents/adapters')
| -rw-r--r-- | collaborativeagents/adapters/personalized_llm_adapter.py | 13 |
1 files changed, 10 insertions, 3 deletions
diff --git a/collaborativeagents/adapters/personalized_llm_adapter.py b/collaborativeagents/adapters/personalized_llm_adapter.py index c2d4727..b476272 100644 --- a/collaborativeagents/adapters/personalized_llm_adapter.py +++ b/collaborativeagents/adapters/personalized_llm_adapter.py @@ -58,9 +58,12 @@ class AdapterConfig: # Best-of-N sampling: generate N responses and pick best (for RAG methods) best_of_n: int = 1 - # Reward mode: "keyword" (legacy heuristic) or "llm" (GPT-5-nano judge) + # Reward mode: "keyword" (legacy heuristic), "llm" (GPT-4o-mini), or "llm_local" (local vLLM) reward_mode: str = "keyword" + # vLLM URL for local reward model (only used when reward_mode="llm_local") + reward_vllm_url: str = "http://localhost:8005/v1" + # Reward mapping for user behavior preference_enforcement_reward: float = -0.8 # Negative reward when user enforces disappointment_expression_reward: float = -0.4 # Milder negative for disappointment @@ -116,6 +119,7 @@ class PersonalizedLLMAdapter: reranker_type=self.config.reranker_type, best_of_n=self.config.best_of_n, reward_mode=self.config.reward_mode, + reward_vllm_url=self.config.reward_vllm_url, ) self._initialized = True print("[Adapter] Initialization complete.") @@ -423,6 +427,7 @@ def create_baseline_adapter( use_vllm: bool = False, use_shared_models: bool = False, reward_mode: str = "keyword", + reward_vllm_url: str = "http://localhost:8005/v1", ) -> PersonalizedLLMAdapter: """ Create an adapter configured for a specific baseline. @@ -438,7 +443,8 @@ def create_baseline_adapter( - "rag_vector": Full personalization (Extractor + RAG + User Vector) device_assignment: GPU assignment dict use_vllm: If True, use vLLM HTTP API for LLM inference (much faster) - reward_mode: Global reward mode ("keyword" or "llm") applied to all methods + reward_mode: Global reward mode ("keyword", "llm", or "llm_local") + reward_vllm_url: vLLM URL for local reward model (when reward_mode="llm_local") use_shared_models: If True, share embedding/reranker models across parallel workers. ESSENTIAL for parallel profile processing to avoid OOM. @@ -592,8 +598,9 @@ def create_baseline_adapter( if device_assignment: config.device_assignment = device_assignment - # Apply global reward_mode to all methods (overrides per-method defaults) + # Apply global reward settings to all methods (overrides per-method defaults) config.reward_mode = reward_mode + config.reward_vllm_url = reward_vllm_url return PersonalizedLLMAdapter(config) |