SitongGong's picture
merge code dirs 1031 (part 3)
4c30587 verified
Raw History Blame Contribute Delete
23.1 kB
"""
Configuration for Agentic Memory System
"""
import argparse
import math
_NEW_ACTION_P_REF = 0.01
def _compute_new_action_delta_max(p_min: float, p_ref: float) -> float:
eps = 1e-8
p_min = min(max(p_min, eps), 1.0 - eps)
p_ref = min(max(p_ref, eps), 1.0 - eps)
return float(math.log((p_min * (1.0 - p_ref)) / (p_ref * (1.0 - p_min))))
class AgenticMemoryConfig:
def __init__(self):
# Memory Bank settings
self.mem_top_k = 5
self.mem_top_k_eval = None
# Operation Bank settings
self.max_ops = 30 # Maximum number of operations in the bank
self.initial_ops = ["insert", "update", "delete", "noop"]
self.skip_noop = False # Skip noop in initial operations
# Controller settings
self.controller_hidden_dim = 256
self.controller_lr = 1e-4
# Encoder models for state and operation embedding
# Default state encoder model supports long sequences up to 4096 tokens
# Alternative: sentence-transformers/all-MiniLM-L6-v2 (faster, shorter context)
self.state_encoder = "allenai/longformer-base-4096"
self.op_encoder = "allenai/longformer-base-4096"
self.use_flash_attn = True # Enable flash attention for compatible encoders
# State fusion strategy for session + memory embeddings
self.state_fusion = "mean" # mean or sim_weighted
self.state_fusion_tau = 1.0 # temperature for sim_weighted
self.encode_batch_size = 64 # batch size for text encoding (to avoid OOM)
# NOTE: state_dim and op_embedding_dim are derived from actual encoder models at runtime
# state_dim = 2 * state_encoder_dim (session_emb || memory_emb concatenated)
# These placeholders will be overwritten in trainer initialization
self.state_dim = None # Will be set from encoder's embedding_dim * 2
self.op_embedding_dim = None # Will be set from encoder's embedding_dim
# Training settings
self.inner_epochs = 20
self.outer_epochs = 10
self.gamma = 0.99 # discount factor
# Episode reward metric: f1 or llm_judge
self.reward_metric = "f1"
# PPO settings
self.gae_lambda = 0.95 # GAE lambda for advantage estimation
self.clip_epsilon = 0.2 # PPO clipping parameter
self.action_top_k = 1 # Number of top actions to select per step
self.vf_clip = 0.0 # value function clipping; 0 disables (recommended default)
self.ppo_epochs = 4 # number of PPO update epochs per batch
self.minibatch_size = 32 # minibatch size for PPO updates; 0 means use full batch
self.entropy_coef = 0.01 # entropy bonus coefficient
self.value_coef = 0.5 # value loss coefficient
# Early stopping: stop PPO epochs when approx_kl exceeds this threshold
# Note: approx_kl is a sample estimate and can be slightly negative due to noise.
self.target_kl = 0.02
self.max_grad_norm = 0.5 # gradient clipping
# Designer settings
self.designer_freq = 1 # trigger every N outer epochs
self.min_op_usage = 5 # minimum usage before considering deletion
self.low_reward_threshold = 0.3
# Case collection settings for designer
self.collect_epochs_before_designer = 5 # Legacy (unused; rolling failure pool is always active)
self.designer_failure_window_epochs = 100 # Rolling window (global inner epochs); 0 disables
self.designer_failure_pool_size = 2000 # Max failures kept in pool; 0 disables
self.designer_num_clusters = 5 # Number of clusters for failure case analysis
self.designer_samples_per_cluster = 3 # Number of cases to sample from each cluster
self.designer_f1_threshold = 0.5 # F1 threshold for success/failure classification
self.op_evolution_trials = 3 # Max retry attempts for operation evolution
self.designer_new_skill_hint = False # Encourage proposing new skills in designer prompts
self.designer_reflection_cycles = 3 # Max reflection cycles for designer analysis
self.designer_max_changes = 1 # Max number of skill changes per evolution cycle
self.designer_refine_only = False # If true, ignore add_new operation changes
# Evolution feedback and early stopping
self.max_designer_evolves = 6 # Maximum number of evolution cycles before stopping
self.designer_early_stop_patience = 3 # Stop if no improvement for N consecutive evolves
self.stage_reward_fraction = 0.25 # Use last 25% of steps in each stage for reward calculation
self.stage_reward_use_moving_avg = False # Use moving average over tail instead of raw mean
self.stage_reward_window = 5 # Window size for moving average if enabled
# Executor settings
self.max_new_tokens = 1024
self.temperature = 0.0
self.designer_model = None
self.llm_judge_model = "openai/gpt-oss-120b"
# Dataset settings
self.dataset = "locomo" # locomo, longmemeval, hotpotqa, or alfworld
self.batch_size = 4 # episodes per PPO update
self.inference_workers = 1 # workers for sample-level memory inference (text datasets)
self.inference_session_workers = 1 # workers for session-level inference within each sample
self.hotpotqa_eval_file = "./data/eval_200.json"
self.alfworld_eval_file = "./data/alfworld_expert_eval_in_distribution.json"
self.locomo_train_query_sampling_ratio = 1.0 # Training-only LoCoMo QA sampling ratio
# HotpotQA specific settings (fixed-length chunking)
self.chunk_size = 1024 # Target chunk size in tokens for fixed-length chunking
self.chunk_overlap = 128 # Overlap between chunks in tokens
# ALFWorld settings (interactive)
self.alfworld_action_max_tokens = 32 # Max tokens for action LLM
self.alfworld_action_temperature = 0.0 # Action LLM temperature
self.alfworld_action_top_p = 1.0 # Action LLM top-p
self.alfworld_include_inventory = True # Include inventory in chunk text
self.alfworld_eval_query_source = "first_observation" # objective or first_observation
# ALFWorld offline pair training settings
self.alfworld_offline_data = None # Path to offline expert trajectories
self.alfworld_pair_a_min = 100 # Min trajectories in batch A
self.alfworld_pair_a_max = 200 # Max trajectories in batch A
self.alfworld_pair_b_size = 0 # Batch B size (interactive envs); 0 uses batch_size
self.alfworld_pair_same_type_prob = 0.8 # Probability A/B share same task type
self.alfworld_pair_chunk_size = 2048 # Token budget for batch-A chunking
self.alfworld_pair_max_steps = 50 # Max steps per interactive episode (batch B)
self.alfworld_pair_b_workers = 64 # Max workers for batch-B envs
self.ema_alpha = 0.1 # EMA smoothing factor for recent usage tracking
# New-action exploration bias (logit adjustment)
self.new_action_p_min = 0.30 # minimum total probability for new actions
# max logit bias applied to new actions, computed for p_ref=0.01
self.new_action_delta_max = _compute_new_action_delta_max(self.new_action_p_min, _NEW_ACTION_P_REF)
self.new_action_bias_steps = 50 # linear decay steps after new actions appear
# Process reward: budget-based allocation for match/mismatch feedback
# Total process reward contribution is bounded by these budgets (episode-level)
self.process_budget = 0.10 # +/- quota per step = process_budget / episode_length
self.failure_budget = 0.05 # penalty for LLM parsing/API failures
self.noop_match_scale = 0.2 # reduce NOOP match reward to discourage "always NOOP" hacking
self.shaping_scale = 1.0 # global scaling factor for shaping rewards
# Reward shaping for long horizon
# redistribute_reward: spread final reward across all steps using exponential decay
# This helps with credit assignment in long episodes
self.redistribute_reward = True
self.reward_redistribution_decay = 0.9 # How much weight decays per step backwards
self.final_reward_last_ratio = 0.6 # Portion added to last step; remainder is redistributed
# Checkpoint loading settings
self.skip_load_operation_bank = False # Skip loading operation bank from checkpoint
self.skip_load_snapshot_manager = False # Skip loading snapshot manager from checkpoint
self.memory_cache_suffix = "" # Optional suffix for memory cache filenames
self.resume_new_wandb_run = False # Resume training from checkpoint but start a fresh wandb run
def update_from_args(self, args):
"""Update config from command line arguments"""
for key, value in vars(args).items():
if hasattr(self, key):
setattr(self, key, value)
if self.mem_top_k_eval is None:
self.mem_top_k_eval = self.mem_top_k
if self.designer_model is None:
self.designer_model = getattr(args, "model", None)
return self
def get_agentic_memory_args():
parser = argparse.ArgumentParser()
# Dataset args
parser.add_argument('--dataset', type=str, default='locomo',
choices=['locomo', 'longmemeval', 'hotpotqa', 'alfworld'])
parser.add_argument('--data-file', type=str, default='./data/locomo10.json')
parser.add_argument('--hotpotqa-eval-file', type=str, default='./data/eval_200.json',
help='HotpotQA eval data file (same format as training)')
parser.add_argument('--alfworld-eval-file', type=str,
default='./data/alfworld_expert_eval_in_distribution.json',
help='ALFWorld eval data file (same format as offline train data)')
parser.add_argument('--locomo-train-query-sampling-ratio', type=float, default=1.0,
help='Training-only LoCoMo QA sampling ratio; 1.0 keeps all non-category-5 queries')
parser.add_argument('--out-file', type=str, default='./results/agentic_memory_results.json')
parser.add_argument('--session-mode', type=str, default='turn-pair',
choices=['turn', 'turn-pair', 'full-session', 'fixed-length', 'paragraph'],
help='Session granularity: turn (each utterance), turn-pair (dialogue exchanges), full-session (entire session), fixed-length (for hotpotqa)')
parser.add_argument('--chunk-size', type=int, default=2048,
help='Target chunk size in tokens for fixed-length chunking (hotpotqa)')
parser.add_argument('--chunk-overlap', type=int, default=256,
help='Overlap between chunks in tokens (hotpotqa)')
# ALFWorld args
parser.add_argument('--alfworld-action-max-tokens', type=int, default=32,
help='Max tokens for ALFWorld action LLM')
parser.add_argument('--alfworld-action-temperature', type=float, default=0.0,
help='Temperature for ALFWorld action LLM')
parser.add_argument('--alfworld-action-top-p', type=float, default=1.0,
help='Top-p for ALFWorld action LLM')
parser.add_argument('--alfworld-include-inventory', type=int, default=1, choices=[0, 1],
help='Include inventory in ALFWorld chunk text (1=yes, 0=no)')
parser.add_argument('--alfworld-eval-query-source', type=str, default='first_observation',
choices=['objective', 'first_observation'],
help='Query source for retrieval in ALFWorld eval')
parser.add_argument('--alfworld-offline-data', type=str, default="./data/alfworld_train_offline.json",
help='Path to offline ALFWorld expert trajectories (pair training)')
parser.add_argument('--alfworld-pair-a-min', type=int, default=40,
help='Min number of trajectories in batch A (offline pair training)')
parser.add_argument('--alfworld-pair-a-max', type=int, default=60,
help='Max number of trajectories in batch A (offline pair training)')
parser.add_argument('--alfworld-pair-b-size', type=int, default=8,
help='Batch B size (interactive envs); 0 uses --batch-size')
parser.add_argument('--alfworld-pair-same-type-prob', type=float, default=1.0,
help='Probability A/B share same task type in offline pair sampling')
parser.add_argument('--alfworld-pair-chunk-size', type=int, default=2048,
help='Token budget for batch-A chunking in offline pair training')
parser.add_argument('--alfworld-pair-max-steps', type=int, default=50,
help='Max steps per interactive episode for batch B')
parser.add_argument('--alfworld-pair-b-workers', type=int, default=64,
help='Max workers for batch-B envs')
parser.add_argument('--overwrite', action='store_true',
help='Force re-extraction of memories even if cached files exist')
# Model args
parser.add_argument('--model', type=str, default='qwen/qwen3-next-80b-a3b-instruct')
parser.add_argument('--designer-model', type=str, default=None,
help='Model for designer LLM; defaults to --model')
parser.add_argument('--api', action='store_true', help='Use API instead of local model')
parser.add_argument('--api-base', type=str, default='[YOUR_API_BASE]')
parser.add_argument('--api-key', type=str, nargs='+',
default=["YOUR_API_KEY_1",
"YOUR_API_KEY_2"])
parser.add_argument('--temperature', type=float, default=0.0)
parser.add_argument('--max-new-tokens', type=int, default=2048)
parser.add_argument('--llm-judge-model', type=str, default='openai/gpt-oss-120b',
help='Model for LLM judge scoring')
parser.add_argument('--batch-size', type=int, default=16, help='Episodes per PPO update')
parser.add_argument('--inference-workers', type=int, default=1,
help='Workers for sample-level memory inference on text datasets (1 = serial)')
parser.add_argument('--inference-session-workers', type=int, default=1,
help='Workers for session-level inference within each sample (1 = serial)')
# Retriever args
parser.add_argument(
'--retriever',
type=str,
default='contriever',
choices=[
'dpr',
'contriever',
'dragon',
'Qwen/Qwen3-Embedding-0.6B',
'qwen/qwen3-embedding-0.6b',
'qwen3-embedding-0.6b',
],
)
parser.add_argument('--mem-top-k', type=int, default=5, help='Top-k memories to retrieve during training')
parser.add_argument('--mem-top-k-eval', type=int, default=None,
help='Top-k memories to retrieve during evaluation (defaults to --mem-top-k)')
parser.add_argument('--max-ops', type=int, default=30,
help='Maximum number of operations in the bank')
parser.add_argument('--skip-noop', action='store_true',
help='Skip noop operation in the initial operation bank')
# Encoder args
parser.add_argument('--state-encoder', type=str, default='Qwen/Qwen3-Embedding-0.6B',
help='Model for state encoding. Use "sentence-transformers/..." or "Qwen/Qwen3-Embedding-0.6B" for SentenceTransformer models')
parser.add_argument('--op-encoder', type=str, default='Qwen/Qwen3-Embedding-0.6B',
help='Model for operation encoding. Use "sentence-transformers/..." or "Qwen/Qwen3-Embedding-0.6B" for SentenceTransformer models')
parser.add_argument('--disable-flash-attn', action='store_false', dest='use_flash_attn', default=True,
help='Disable flash attention for compatible embedding encoders')
parser.add_argument('--state-fusion', type=str, default='mean',
choices=['mean', 'sim_weighted'],
help='Fusion strategy for session and memory embeddings')
parser.add_argument('--state-fusion-tau', type=float, default=1.0,
help='Temperature for sim_weighted fusion (higher -> flatter weights)')
parser.add_argument('--encode-batch-size', type=int, default=8,
help='Batch size for text encoding (to avoid OOM on large batches)')
# Training args
parser.add_argument('--inner-epochs', type=int, default=100)
parser.add_argument('--outer-epochs', type=int, default=10)
parser.add_argument('--controller-lr', type=float, default=1e-4)
parser.add_argument('--gamma', type=float, default=0.99, help='Discount factor')
parser.add_argument('--reward-metric', type=str, default='f1',
choices=['f1', 'llm_judge'],
help='Episode reward metric: f1 or llm_judge')
# PPO args
parser.add_argument('--gae-lambda', type=float, default=0.95, help='GAE lambda for advantage estimation')
parser.add_argument('--clip-epsilon', type=float, default=0.2, help='PPO clipping parameter')
parser.add_argument('--action-top-k', type=int, default=1,
help='Number of top actions to select per step (default 1)')
parser.add_argument('--vf-clip', type=float, default=0.0,
help='Value function clipping range; 0 disables value clipping')
parser.add_argument('--ppo-epochs', type=int, default=4, help='Number of PPO update epochs per batch')
parser.add_argument('--minibatch-size', type=int, default=32, help='Minibatch size for PPO; 0 uses full batch')
parser.add_argument('--entropy-coef', type=float, default=0.01, help='Entropy bonus coefficient')
parser.add_argument('--value-coef', type=float, default=0.5, help='Value loss coefficient')
parser.add_argument('--target-kl', type=float, default=0.02,
help='Early stop PPO update epochs when approx_kl exceeds this threshold')
parser.add_argument('--max-grad-norm', type=float, default=0.5, help='Gradient clipping')
parser.add_argument('--new-action-p-min', type=float, default=0.30,
help='Minimum total probability mass for new actions')
parser.add_argument('--new-action-bias-steps', type=int, default=50,
help='Linear decay steps for new-action bias after changes')
# Designer args
parser.add_argument('--designer-freq', type=int, default=1, help='Trigger designer every N outer epochs')
parser.add_argument('--enable-designer', action='store_true', help='Enable operation evolution')
parser.add_argument('--collect-epochs-before-designer', type=int, default=5,
help='Legacy (unused; rolling failure pool is always active)')
parser.add_argument('--designer-failure-window-epochs', type=int, default=100,
help='Rolling window in global inner epochs for failure pool; 0 disables')
parser.add_argument('--designer-failure-pool-size', type=int, default=2000,
help='Max failures kept in pool; 0 disables')
parser.add_argument('--designer-num-clusters', type=int, default=5,
help='Number of clusters for failure case analysis')
parser.add_argument('--designer-samples-per-cluster', type=int, default=3,
help='Number of cases to sample from each cluster')
parser.add_argument('--designer-f1-threshold', type=float, default=0.5,
help='F1 threshold for success/failure classification')
parser.add_argument('--designer-new-skill-hint', action='store_true',
help='Encourage proposing new skills in designer prompts')
parser.add_argument('--designer-reflection-cycles', type=int, default=3,
help='Max reflection cycles for designer analysis (>=1)')
parser.add_argument('--designer-max-changes', type=int, default=1,
help='Max number of skill changes per evolution cycle (>=1)')
parser.add_argument('--designer-refine-only', action='store_true',
help='Only allow refining existing skills; ignore add_new changes')
parser.add_argument('--op-evolution-trials', type=int, default=3,
help='Max retry attempts for operation evolution')
parser.add_argument('--max-designer-evolves', type=int, default=20,
help='Maximum number of evolution cycles before stopping training')
parser.add_argument('--designer-early-stop-patience', type=int, default=5,
help='Stop training if no improvement for N consecutive evolves')
parser.add_argument('--stage-reward-fraction', type=float, default=0.25,
help='Use last X fraction of steps in each stage for reward calculation')
parser.add_argument('--stage-reward-use-moving-avg', action='store_true',
help='Use moving average over tail instead of raw mean for stage reward')
parser.add_argument('--stage-reward-window', type=int, default=5,
help='Window size for moving average stage reward (>=1)')
# Misc
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--device', type=str, default='cuda')
parser.add_argument('--save-dir', type=str, default='./checkpoints')
parser.add_argument('--log-dir', type=str, default='./logs')
parser.add_argument('--round', type=int, default=10, help='Number of retry rounds for LLM API calls')
# Evaluation
parser.add_argument('--eval-only', action='store_true')
parser.add_argument('--load-checkpoint', type=str, default=None)
parser.add_argument('--skip-load-operation-bank', action='store_true',
help='Skip loading operation bank from checkpoint')
parser.add_argument('--skip-load-snapshot-manager', action='store_true',
help='Skip loading snapshot manager from checkpoint')
parser.add_argument('--memory-cache-suffix', type=str, default='',
help='Optional suffix appended to memory cache filenames')
# Wandb args
parser.add_argument('--wandb-project', type=str, default='memskill', help='Wandb project name')
parser.add_argument('--wandb-run-name', type=str, default=None, help='Wandb run name')
parser.add_argument('--wandb-key', type=str, default=None,
help='Wandb API key (optional; can also use WANDB_API_KEY env var)')
parser.add_argument('--resume-new-wandb-run', action='store_true',
help='When resuming from a checkpoint, start a new wandb run instead of resuming the saved run ID')
args = parser.parse_args()
return args