config.yaml

# MUD World Model — runtime and training config

# MUD connection
mud:
  silence_timeout_sec: 10
  reconnect_delay_sec: 2
  max_reconnect_attempts: 5

# Paths (relative to project root or absolute)
paths:
  data_dir: data
  memory_dir: data          # commands.md, spells.md, current_location.md, mobs.md, session_summary.md, goals.md, inventory.md, equipment.md, statbar.md
  logs_dir: data/logs
  traces_file: data/logs/traces.jsonl
  training_data_dir: data/training_data   # archived traces and orchestrator logs (flat timestamps); training uses trace_glob over these
  checkpoints_dir: data/checkpoints/wm
  # wm_checkpoint_run: run_20260203_120000   # optional: pin WM to this run for rollback; if unset, latest run_* is used
  prompts_dir: prompts

# API agents (OpenAI)
openai:
  model: gpt-5-nano          # default fallback
  model_ph: gpt-4o-mini      # PH: faster runtime
  model_mh: gpt-4o-mini      # MH: faster, sufficient for memory parsing
  model_dh: gpt-4o-mini      # DH: faster, sufficient for choosing from options
  model_vh: gpt-4o-mini      # VH: faster; can compensate with prompting/memory
  temperature: 1   # gpt-5-nano only supports 1; omit or set lower for models that allow it (e.g. gpt-4o-mini)

# World model (local Mistral-7B)
wm:
  model_name: mistralai/Mistral-7B-v0.1
  use_4bit: false           # set true if low VRAM
  max_new_tokens: 256
  temperature: 0.3

# Orchestrator
orchestrator:
  monitor_interval: 10      # every N turns, monitor agent checks if player is stuck and may teleport
  kickoff_commands:
    - look
    - score
    - inventory
    - equipment
    - practice   # when not at guildmaster, shows known spells/skills
  max_steps: null           # null = run until stopped
  game_buffer_max_lines: 50   # DH sees this many lines; PH uses ph_buffer_lines for its call
  ph_buffer_lines: 25         # PH sees this many lines of game buffer (reduces PH input size)
  wm_buffer_lines: 25         # WM sees this many lines of game buffer for next-line context (local model)

# Training
training:
  trace_glob: data/training_data/traces_*.jsonl   # cumulative archived runs; each main.py run appends traces_YYYYMMDD_HHMMSS.jsonl
  vh_score_min: 1           # keep all traces; vh_score rates prediction quality, not outcome (see train.build_dataset)
  weight_by_vh_score: true
  mode: outcome_summary     # "next_line" or "outcome_summary"
  lora_r: 8
  lora_alpha: 16
  max_length: 2048                # max sequence length (context + target). Fit composed context + 25-line buffer + action + target; Mistral supports 8192.
  per_device_train_batch_size: 2   # 512-token seqs: 24GB VRAM ~2–4; 16GB ~1–2. Use --batch-size 1 if OOM.
  gradient_accumulation_steps: 4   # effective batch = per_device * this (default 8)
  num_epochs: 25
  learning_rate: 2.0e-5
  early_stopping_patience: 2   # stop after this many epochs with no improvement on eval loss
  eval_fraction: 0.1           # fraction of data used for validation (0 = no early stopping)
  # Semantic loss (Option A): match WM representation to VH summary in embedding space (trains meaning, not wording)
  use_semantic_loss: true
  semantic_encoder: sentence-transformers/all-MiniLM-L6-v2
  semantic_loss_weight: 0.5   # L = L_ce + this * L_sem (hybrid). Set semantic_only: true to use only L_sem.
  semantic_only: false