config.example.json

{
  "_comment": "ForgeLoopGRPO Config — Supports Qwen3.5-4B and Gemma-4-E4B-IT",
  "_comment_model": "Set model_path to 'Qwen/Qwen3.5-4B' or 'google/gemma-4-E4B-it'",

  "model_path": "Qwen/Qwen3.5-4B",
  "dataset_path": "data/dataset.jsonl",
  "output_dir": "outputs/forge_loop_grpo",
  "num_generations": 8,
  "per_device_train_batch_size": 1,
  "gradient_accumulation_steps": 8,
  "beta": 0.04,
  "learning_rate": 5e-06,
  "temperature": 0.9,
  "top_p": 0.95,
  "max_new_tokens": 2048,
  "max_seq_length": 4096,
  "num_train_epochs": 3,
  "reward_weights": {
    "thematic_consistency": {"enabled": true, "weight": 0.15},
    "tone_consistency": {"enabled": true, "weight": 0.1},
    "semantic_diversity": {"enabled": true, "weight": 0.1},
    "fluency": {"enabled": true, "weight": 0.1},
    "reasoning_depth": {"enabled": true, "weight": 0.05},
    "lexical_diversity": {"enabled": true, "weight": 0.05},
    "efficiency_coefficient": {"enabled": true, "weight": 0.05},
    "directive_clarity": {"enabled": true, "weight": 0.05},
    "context_alignment": {"enabled": true, "weight": 0.05},
    "input_adaptation": {"enabled": true, "weight": 0.05},
    "style_preservation": {"enabled": true, "weight": 0.05},
    "creative_problem_solving": {"enabled": true, "weight": 0.03},
    "cognitive_richness": {"enabled": true, "weight": 0.03},
    "style_coherence": {"enabled": true, "weight": 0.03},
    "exploratory_boldness": {"enabled": true, "weight": 0.03}
  },
  "soft_cull": {
    "enabled": true,
    "cull_percentage": 0.03,
    "cull_penalty": -0.05
  },
  "degeneracy_auditor": {
    "severe_looping": {"enabled": true, "penalty": -0.1},
    "missing_think_tag": {"enabled": true, "penalty": -0.1},
    "linguistic_explosion": {"enabled": true, "penalty": -0.1}
  },
  "think_tag_start": "<think>",
  "think_tag_end": "</think>",
  "gemma_thinking_enabled": false,
  "gate": {
    "enabled": true,
    "warmup_steps": 0,
    "historical_mean_alpha": 0.1
  },
  "vllm": {
    "tensor_parallel_size": 1,
    "gpu_memory_utilization": 0.50,
    "max_num_seqs": 47,
    "max_model_len": 4096,
    "dtype": "auto",
    "swap_space": 2
  },
  "performance": {
    "embedding_device": "cpu",
    "embedding_model": "all-MiniLM-L6-v2"
  },
  "gradient_checkpointing": true,
  "lora": {
    "r": 64,
    "alpha": 128,
    "dropout": 0.05,
    "target_modules": [
      "q_proj", "k_proj", "v_proj", "o_proj",
      "gate_proj", "up_proj", "down_proj"
    ]
  },
  "save_steps": 500,
  "eval_steps": 500,
  "semantic_diversity_floor": 0.01,
  "torch_compile": false,
  "auto_merge_final": true,
  "domain_constants": {
    "vocab": [
      "analysis", "synthesis", "evaluation", "conclusion", "hypothesis",
      "evidence", "reasoning", "insight", "perspective", "framework",
      "clarity", "truth", "understanding", "discovery", "inquiry",
      "method", "principle", "theory", "concept", "abstraction",
      "context", "nuance", "implication", "interpretation", "reflection",
      "observation", "inference", "deduction", "induction", "validation",
      "critique", "examination", "exploration", "investigation", "contemplation",
      "deliberation", "judgment", "wisdom", "knowledge", "learning",
      "growth", "development", "progress", "evolution", "transformation",
      "adaptation", "resilience", "integrity", "purpose", "meaning"
    ],
    "anti_patterns": [
      "as an ai", "i am an ai", "language model", "llm",
      "artificial intelligence", "i cannot", "i'm sorry", "i apologize",
      "i cannot fulfill", "i'm not able", "i don't have", "i have no",
      "i cannot provide", "i cannot assist"
    ],
    "sentiment_markers": {
      "positive": [
        "excellent", "efficient", "optimal", "beneficial", "valuable",
        "effective", "successful"
      ],
      "negative": [
        "error", "failure", "flaw", "inefficient", "deficient",
        "suboptimal", "incorrect"
      ],
      "neutral": [
        "objective", "standard", "consistent", "uniform",
        "stable", "balanced", "systematic"
      ]
    },
    "action_verbs": [
      "execute", "implement", "analyze", "evaluate", "modify",
      "generate", "process", "maintain", "secure", "optimize",
      "compile", "verify", "integrate", "deploy", "configure",
      "validate", "structure", "direct", "lead", "guide"
    ],
    "rejection_markers": [
      "cannot fulfill", "is unable to", "not supported",
      "invalid request", "restricted", "outside the scope",
      "terminated", "does not comply"
    ],
    "thematic_markers": [
      "domain", "context", "framework", "structure", "core",
      "pivotal", "fundamental", "origin", "basis", "background",
      "scope", "influence", "parameter"
    ],
    "reasoning_markers": [
      "therefore", "because", "thus", "hence", "consequently",
      "furthermore", "additionally", "alternatively", "whereas",
      "implicitly", "explicitly", "specifically", "in contrast"
    ],
    "absolute_markers": [
      "must", "will", "shall", "cannot", "never", "always", "absolutely"
    ],
    "creative_markers": [
      "imagine", "picture", "like a", "similar to", "analogy",
      "what if", "reverse", "backwards", "different way", "another angle"
    ],
    "cognitive_modalities": {
      "math": ["calculate", "equals", "plus", "minus", "sum", "ratio", "number"],
      "logical": ["premise", "syllogism", "valid", "invalid", "fallacy", "assertion"],
      "strategic": ["plan", "strategy", "tactic", "move", "position", "advantage"],
      "empirical": ["observe", "data", "metric", "evidence", "test", "experiment"]
    },
    "structural_phrases": [
      "initially", "firstly", "consequently", "subsequently",
      "furthermore", "in conclusion", "summarizing", "moving forward",
      "at this stage", "ultimately", "finally"
    ]
  },
  "chat_template": {
    "im_start": "<|im_start|>",
    "im_end": "<|im_end|>",
    "use_system_prompt": true,
    "system_prompt_steps": 200,
    "system_prompt_text": "You are a helpful, precise, and logically sound assistant focused on clear text generation and accurate instruction following."
  }
}