config/learning.yaml

# Continuous learning configuration for Jekyll & Hyde

paths:
  interactions: data/learning/interactions.jsonl
  curated: data/learning/curated_train.jsonl
  rejected: data/learning/rejected.jsonl
  state: data/learning/state.json

quality:
  min_assistant_chars: 80
  min_quality_score: 0.65
  require_markdown: true
  reject_leak_patterns:
    - "^model$"
    - "<start_of_turn>"
    - "Hyde test probe:"
    - "RESPONSE TEMPLATE"
    - "KEY CONCEPT"
    - "SAMPLE ANSWER"
    - "USER QUERY:"
    - "Policy Analyst Role"

auto:
  curate_after_each_turn: true
  merge_dataset_on_curate: false
  train_when_curated_reaches: 20
  train_min_interval_hours: 6
  train_epochs_incremental: 2
  train_base: gemma2-2b
  auto_train_enabled: true
  adapter_only: true

feedback:
  upvote_boost: 0.35
  downvote_reject: true

diet:
  enabled: true
  semantic_threshold: 0.92
  embedding_model: BAAI/bge-small-en-v1.5
  embedding_index: data/learning/embedding_index.json
  max_total_records: 2000
  category_caps:
    quant: 400
    policy: 500
    duel: 300
    chat: 800
    mcp_tools: 250
  persona_caps:
    jekyll: 800
    hyde: 600
    neutral: 1200

quantize:
  export_gguf_after_train: true
  gguf_quant: q4_k_m
  prune_old_gguf: true

# Gray-zone reinforcement: duel → extract zones → dual synthesis → curated training
gray_reinforce:
  enabled: true
  synthesize_solutions: true
  use_llm_synthesis: true
  auto_curate: true
  min_zones: 1
  max_zones_per_duel: 8

# RLAIF gate — verification API score before auto-curation (prevents hallucination loops)
rlaif:
  enabled: true
  min_score: 85

# Distilled gray-zone rule memory (RAG injection + eviction preservation)
memory:
  enabled: true
  inject_on_query: true
  distill_on_evict: true
  max_entries: 500
  retrieve_k: 3
  min_similarity: 0.55
  consolidate_after: 40
  consolidate_similarity: 0.88
  consolidation_enabled: true
  paths:
    rules: data/learning/memory_rules.jsonl

# LoRA MoE — dynamic jekyll/hyde adapter blend
lora_moe:
  enabled: true
  quantize_step: 0.05
  bucket_cache: true
  buckets:
    - { jekyll: 0.9, hyde: 0.1 }
    - { jekyll: 0.7, hyde: 0.3 }
    - { jekyll: 0.5, hyde: 0.5 }
    - { jekyll: 0.3, hyde: 0.7 }
    - { jekyll: 0.1, hyde: 0.9 }

# MCP structured tool-calling training bucket (high priority in data diet)
mcp_training:
  enabled: true
  priority_boost: 0.35

# Dynamic decoding — MoE blend ratio → temperature / top-p / min-p
decoding:
  dynamic_entropy: true
  link_to_moe: true
  jekyll_temperature: 0.15
  jekyll_top_p: 0.85
  jekyll_min_p: 0.08
  hyde_temperature: 0.35
  hyde_top_p: 0.92
  hyde_min_p: 0.05
  blend_temp_min: 0.2
  blend_temp_max: 0.78
  blend_top_p_min: 0.88
  blend_min_p_min: 0.06

# Grammar-constrained MCP JSON tool output at decode time
grammar:
  mcp_tool_json: true

# DPO preference alignment from curated vs rejected (RLAIF pairs)
dpo:
  enabled: true
  prefer_over_sft: true
  min_pairs: 2
  beta: 0.1
  epochs: 1
  dataset_path: data/learning/dpo_pairs.jsonl