Module 11: trained all-layer LoRA and matched replay

{
  "module": 11,
  "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
  "revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
  "steps": 200,
  "selected_step": 200,
  "trainable_parameters": 460800,
  "adapter_update_norm": 5.6753139172183555,
  "adapter_replay_verified": true,
  "validation_loss_before": 1.535772931575775,
  "validation_loss_after": 0.0030356942541402534,
  "validation_base_schema_rate": 0.0,
  "validation_adapted_schema_rate": 1.0,
  "base_test_completion_loss": 1.531955510377884,
  "adapted_test_completion_loss": 0.029949328132061054,
  "base_generation_accuracy": 0.0,
  "adapted_generation_accuracy": 0.94,
  "base_schema_rate": 0.0,
  "adapted_schema_rate": 1.0,
  "test_examples": 100,
  "rate_intervals": [
    {
      "method": "base",
      "metric": "schema_validity",
      "successes": 0,
      "observations": 100,
      "rate": 0.0,
      "ci_low": 0.0,
      "ci_high": 0.03699349820698566,
      "confidence": 0.95
    },
    {
      "method": "base",
      "metric": "accuracy",
      "successes": 0,
      "observations": 100,
      "rate": 0.0,
      "ci_low": 0.0,
      "ci_high": 0.03699349820698566,
      "confidence": 0.95
    },
    {
      "method": "adapted",
      "metric": "schema_validity",
      "successes": 100,
      "observations": 100,
      "rate": 1.0,
      "ci_low": 0.9630065017930143,
      "ci_high": 1.0,
      "confidence": 0.95
    },
    {
      "method": "adapted",
      "metric": "accuracy",
      "successes": 94,
      "observations": 100,
      "rate": 0.94,
      "ci_low": 0.8752318455410406,
      "ci_high": 0.9722138760368118,
      "confidence": 0.95
    }
  ],
  "quality_claim": "Historical class-balanced sample; format compliance and label correctness are separate"
}

Only all-layer q/v LoRA parameters train. The cached base remains unchanged; the best validation-loss checkpoint is loaded before final evaluation.

Training/validation/final subsets contain disjoint real text IDs. Prompt and padding tokens carry label -100.

Schema-valid JSON can still have the wrong label. The balanced final subset changes prevalence; binomial intervals do not correct sampling bias.

Adapters are saved as safetensors with PEFT configuration and exact base revision; reloaded logits must match.