code/self_healing_finetuning/model_policy/roadmap.json

{
  "created_at": "2026-05-14",
  "policy_version": "model_roadmap_v1",
  "strategy": {
    "current_default": "fingpt_first_bootstrap",
    "current_default_model": "linvest21/linvest21_fingpt_v1_000",
    "source_adapter": "FinGPT/fingpt-mt_llama3-8b_lora",
    "source_base_model": "meta-llama/Meta-Llama-3-8B",
    "rationale": [
      "Use the newest official FinGPT multitask Llama3 adapter as the Linvest21 bootstrap.",
      "Do not repeat broad finance-domain training already represented by FinGPT.",
      "Use SHFT cycles to add Linvest21-specific corpus, evaluation, self-healing, and promotion evidence.",
      "Preserve a foundation-model upgrade path for long-context and higher-capability bases."
    ]
  },
  "quantitative_decision_weights": {
    "base_capability": 0.3,
    "finance_bootstrap": 0.25,
    "context_window": 0.2,
    "trainability": 0.15,
    "license_and_operational_readiness": 0.1
  },
  "candidates": [
    {
      "id": "linvest21/linvest21_fingpt_v1_000",
      "track": "current_fingpt_bootstrap",
      "source_adapter": "FinGPT/fingpt-mt_llama3-8b_lora",
      "base_model": "meta-llama/Meta-Llama-3-8B",
      "parameter_billions": 8,
      "context_tokens": 8192,
      "finance_bootstrap_score": 1.0,
      "base_capability_score": 0.7,
      "trainability_score": 0.95,
      "license_and_operational_readiness_score": 0.85,
      "recommended_use": "default SHFT bootstrap and near-term Linvest21 v1 iteration base",
      "gates": {
        "hf_repo_created": true,
        "provenance_recorded": true,
        "finetune_ready": true,
        "long_context_ready": false
      }
    },
    {
      "id": "FinGPT/fingpt-sentiment_llama2-13b_lora",
      "track": "legacy_task_specialist",
      "source_adapter": "FinGPT/fingpt-sentiment_llama2-13b_lora",
      "base_model": "meta-llama/Llama-2-13b",
      "parameter_billions": 13,
      "context_tokens": 4096,
      "finance_bootstrap_score": 0.8,
      "base_capability_score": 0.55,
      "trainability_score": 0.85,
      "license_and_operational_readiness_score": 0.7,
      "recommended_use": "sentiment benchmark/reference only, not Linvest21 default base",
      "gates": {
        "hf_repo_created": false,
        "provenance_recorded": false,
        "finetune_ready": true,
        "long_context_ready": false
      }
    },
    {
      "id": "FinGPT/fingpt-mt_chatglm2-6b_lora",
      "track": "legacy_longer_context_fingpt",
      "source_adapter": "FinGPT/fingpt-mt_chatglm2-6b_lora",
      "base_model": "THUDM/chatglm2-6b",
      "parameter_billions": 6,
      "context_tokens": 32768,
      "finance_bootstrap_score": 0.75,
      "base_capability_score": 0.45,
      "trainability_score": 0.75,
      "license_and_operational_readiness_score": 0.65,
      "recommended_use": "context-window comparison and Chinese-market reference, not default Linvest21 base",
      "gates": {
        "hf_repo_created": false,
        "provenance_recorded": false,
        "finetune_ready": true,
        "long_context_ready": true
      }
    },
    {
      "id": "meta-llama/Llama-3.1-8B-Instruct",
      "track": "foundation_upgrade_candidate",
      "source_adapter": null,
      "base_model": "meta-llama/Llama-3.1-8B-Instruct",
      "parameter_billions": 8,
      "context_tokens": 131072,
      "finance_bootstrap_score": 0.2,
      "base_capability_score": 0.82,
      "trainability_score": 0.9,
      "license_and_operational_readiness_score": 0.65,
      "recommended_use": "first long-context foundation upgrade candidate after Linvest21 eval gates exist",
      "gates": {
        "hf_repo_created": false,
        "provenance_recorded": false,
        "finetune_ready": true,
        "long_context_ready": true
      }
    },
    {
      "id": "meta-llama/Llama-3.1-70B-Instruct",
      "track": "foundation_upgrade_candidate",
      "source_adapter": null,
      "base_model": "meta-llama/Llama-3.1-70B-Instruct",
      "parameter_billions": 70,
      "context_tokens": 131072,
      "finance_bootstrap_score": 0.2,
      "base_capability_score": 0.92,
      "trainability_score": 0.55,
      "license_and_operational_readiness_score": 0.55,
      "recommended_use": "higher-capability long-context target when budget and hardware permit",
      "gates": {
        "hf_repo_created": false,
        "provenance_recorded": false,
        "finetune_ready": true,
        "long_context_ready": true
      }
    },
    {
      "id": "Qwen/Qwen2.5-7B-Instruct",
      "track": "foundation_upgrade_candidate",
      "source_adapter": null,
      "base_model": "Qwen/Qwen2.5-7B-Instruct",
      "parameter_billions": 7,
      "context_tokens": 131072,
      "finance_bootstrap_score": 0.1,
      "base_capability_score": 0.78,
      "trainability_score": 0.9,
      "license_and_operational_readiness_score": 0.8,
      "recommended_use": "non-Llama long-context comparison candidate",
      "gates": {
        "hf_repo_created": false,
        "provenance_recorded": false,
        "finetune_ready": true,
        "long_context_ready": true
      }
    }
  ],
  "promotion_rules": {
    "stay_on_fingpt_bootstrap_until": [
      "Step 0 learning corpus contains only valid PDF, curated JSONL, or normalized policy-approved source packages and passes audit.",
      "Baseline eval exists for linvest21/linvest21_fingpt_v1_000.",
      "At least one SHFT train/eval/repair cycle produces candidate evidence."
    ],
    "allow_foundation_upgrade_when": [
      "Candidate beats current Linvest21 FinGPT baseline on aggregate eval by at least 0.5 percent relative and no critical task regression exceeds 2 percent absolute.",
      "Candidate supports at least 128K context or demonstrates materially better reasoning at the same context.",
      "License and gated-model access are approved.",
      "Serving cost is within the configured budget cap."
    ]
  }
}