20260628-qwen-small-logits-orchestrator-split-smoke-plan.json

{
  "can_train_here": true,
  "config": {
    "backend": "transformers",
    "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
    "batch_size": 1,
    "epochs": 1,
    "learning_rate": 0.0002,
    "lora_rank": 0,
    "max_length": 512,
    "seed": 7,
    "train_backbone": false
  },
  "dependency_status": {
    "mlx": false,
    "mlx_lm": false,
    "torch": true,
    "transformers": true
  },
  "prepared_rows": "research/datasets/20260628-qwen-small-logits-orchestrator-split-train.jsonl",
  "record_count": 53,
  "rows_output": "research/datasets/20260628-qwen-small-logits-orchestrator-split-train.jsonl",
  "schema_version": "mempool.qwen_logits_orchestrator_plan.v1",
  "training_order": [
    "freeze Qwen-small backbone",
    "train worker/workflow/verifier/abstain heads on measured soft targets",
    "compare held-out routing against the linear multi-head baseline",
    "only enable LoRA/backbone updates after the heads beat the baseline"
  ],
  "worker_ids": [
    "ollama-cloud-deepseek-v4-pro",
    "ollama-cloud-glm-5.2",
    "ollama-cloud-kimi-k2.7-code",
    "ollama-cloud-qwen3-coder-480b"
  ],
  "workflow_labels": [
    "direct",
    "verify_then_fallback"
  ]
}