Trained / exported adapter / adapter-e4b_opd_r32_v1_offpolicy_overfit

e4b_opd_r32_v1_offpolicy_overfit

gemma-4-e4b-it-qat-w4a16-ctAuthor: Not recorded in adapter manifestState: Archived checkpointVerdict: Historical measurements

When

Recorded date: 2026-11-01

First archive commit: 2026-11-04T17:35:03-05:00 (offset shown).

Training run day: 2026-11-01 (America/New_York). For derived variants, this is the parent run day, not the variant’s creation day.

First Git archive commit time (UTC offset shown), not the time training changed weights. Per-update wall timestamps were not preserved.

Exact tensor-change timestamp: unavailable.

Why change it?

Transfer successful repairs generated by the larger 31B model into E4B.

Which part changed?

The source describes adapter targets and mathematical changes. No before/after tensor comparison is available in this viewer.

How was it changed?

Off-policy prediction distillation from 26 teacher-solved trajectories across 11 tasks; 3 epochs, 20 optimizer steps. Historical ledger alias: e4b_opd_r32_v1_offpolicy. Exported low-rank adapter, rank 32, scaling alpha 64.

Did it improve?

Historical measurements · Historical local result: 3/15 on val15.

An exported checkpoint is not proof of a promoted policy or a complete optimizer-step history.

Evaluation set: Historical sets vary; see source table · Batch: Not applicable. Local screening and historical task scores are not competition scores unless explicitly identified in the source. Different evaluation sets cannot be compared as a gain.

Original evidence

Literal source excerpts follow. They may contain original predictions and superseded counts; use the measured-outcome summary and corrections above for the current result.

Adapter configuration — original metadata

{
  "base_model_name_or_path": "gemma-4-e4b-it-qat-w4a16-ct",
  "bias": "none",
  "lora_alpha": 64,
  "lora_dropout": 0.0,
  "peft_type": "LORA",
  "r": 32,
  "target_modules": "^model\\.language_model\\.layers\\.\\d+\\.(self_attn\\.(q_proj|k_proj|v_proj|o_proj)|mlp\\.(gate_proj|up_proj|down_proj))$",
  "task_type": "CAUSAL_LM",
  "use_rslora": false
}

Checkpoint identity — original manifest

{
  "adapter_id": "e4b_opd_r32_v1_offpolicy_overfit",
  "wrx90_path": "/home/keith/gemma-v2/adapters/e4b_opd_r32_v1_offpolicy_overfit",
  "files": {
    "adapter_model.safetensors": {
      "size_bytes": 139603296,
      "sha256": "ab3fbc766aabc8d80f0439f4765dcd8676871ddcc7847a7a675a6cbcde0ba0c0"
    }
  }
}

Training metrics — original archive, not task improvement

{
  "dataset_size": 26,
  "elapsed_seconds": 3420.96,
  "epochs": 3,
  "export": {
    "adapter_dir": "/home/keith/gemma-v2/adapters/e4b_opd_r32_v1",
    "safetensors_bytes": 139603296,
    "tensor_count": 516
  },
  "history_head": [
    {
      "ce_loss": 0.3524,
      "epoch": 1,
      "grad_norm": 0.5938,
      "loss": 0.4286,
      "lr_a": 7.5e-06,
      "opt_step": 1,
      "rkl_loss": 0.5049
    },
    {
      "ce_loss": 0.4436,
      "epoch": 1,
      "grad_norm": 0.6641,
      "loss": 0.5385,
      "lr_a": 1.5e-05,
      "opt_step": 2,
      "rkl_loss": 0.6335
    },
    {
      "ce_loss": 0.5266,
      "epoch": 1,
      "grad_norm": 0.6289,
      "loss": 0.6417,
      "lr_a": 1.5e-05,
      "opt_step": 3,
      "rkl_loss": 0.7567
    }
  ],
  "history_tail": [
    {
      "ce_loss": 0.3167,
      "epoch": 3,
      "grad_norm": 0.21,
      "loss": 0.3353,
      "lr_a": 5.54e-06,
      "opt_step": 15,
      "rkl_loss": 0.354
    },
    {
      "ce_loss": 0.2918,
      "epoch": 3,
      "grad_norm": 0.5195,
      "loss": 0.2899,
      "lr_a": 4.56e-06,
      "opt_step": 16,
      "rkl_loss": 0.288
    },
    {
      "ce_loss": 0.4374,
      "epoch": 3,
      "grad_norm": 0.1875,
      "loss": 0.4399,
      "lr_a": 3.68e-06,
      "opt_step": 17,
      "rkl_loss": 0.4424
    },
    {
      "ce_loss": 0.3092,
      "epoch": 3,
      "grad_norm": 0.2461,
      "loss": 0.3177,
      "lr_a": 2.92e-06,
      "opt_step": 18,
      "rkl_loss": 0.3262
    },
    {
      "ce_loss": 0.2636,
      "epoch": 3,
      "grad_norm": 0.2031,
      "loss": 0.2551,
      "lr_a": 2.31e-06,
      "opt_step": 19,
      "rkl_loss": 0.2467
    }
  ],
  "injected_module_count": 258,
  "mode": "dual_gpu_opd",
  "opt_steps": 20,
  "student_alias": "gemma-4-e4b-it-qat-w4a16-ct"
}