curriculum_stages.json

{
  "ablations": [
    "sft_k3_only",
    "sft_qwen_v4_supplement",
    "sft_glm_agent"
  ],
  "default_entrypoint": "sft_balanced",
  "schema_version": "sota-curriculum-v1",
  "specialists": [
    "sft_math",
    "sft_code",
    "sft_science",
    "sft_long_context",
    "sft_multilingual",
    "sft_dialogue",
    "sft_instruction_final"
  ],
  "stages": [
    {
      "config": "sft_balanced",
      "epochs": 1,
      "id": "A_generalist",
      "notes": "Default generalist entrypoint with rebalanced sampling weights"
    },
    {
      "config": "sft_tools",
      "epochs": 0.5,
      "id": "B_tools",
      "notes": "Tool/agent specialization",
      "optional_mix": [
        "sft_agent"
      ]
    },
    {
      "config": "sft_reasoning_specialist",
      "epochs": 0.5,
      "id": "C_reasoning_optional",
      "lr_note": "Use lower LR than stage A",
      "notes": "Optional reasoning push; dataset already reasoning-heavy"
    },
    {
      "config": "rl_tool_prompts",
      "id": "D_rl_tools_optional",
      "method": "GRPO/RLVR",
      "notes": "Verifiable tool tasks only; not for open-ended dialogue"
    }
  ]
}