Trained / exported adapter / adapter-e4b_true_opd_v5

e4b_true_opd_v5

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

When

Recorded date: 2026-11-02

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

Training run day: 2026-11-02 (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?

Add negative repair examples to online prediction distillation.

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?

Online prediction distillation with the recorded CISPO-negative objective; 116 trajectories (95 resolved, 21 negative), 2 epochs, 58 optimizer steps. Exported low-rank adapter, rank 64, scaling alpha 32.

Did it improve?

Historical measurements · Historical local result: 8/45.

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 They may contain original predictions and superseded counts; use the measured-outcome summary and corrections above for the current result.

Historical training and evaluation ledger — original table rows

| 6 | `e4b_true_opd_v5` | Nov 2 | Online OPD + CISPO Negatives | `E4B` `r=64, a=32` | `opd_e4b_cispo_v5` | `116` (`95` res + `21` neg, `194` tasks) | `0.57M` | `2` (`58`) | `1.14M` | `1.05h` | `8/45` (`14` errs; WiSE-FT `7/45`) | — |

Adapter configuration — original metadata

{
  "base_model_name_or_path": "gemma-4-e4b-it-qat-w4a16-ct",
  "bias": "none",
  "lora_alpha": 32,
  "lora_dropout": 0.0,
  "peft_type": "LORA",
  "r": 64,
  "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_true_opd_v5",
  "wrx90_path": "/home/keith/gemma-v2/adapters/e4b_true_opd_v5",
  "files": {
    "adapter_model.safetensors": {
      "size_bytes": 279129832,
      "sha256": "8a38dd05402aeb956299fe2e66765860b8a604ac4e27ceacbf4769be923d2fe1"
    }
  }
}

Training metrics — original archive, not task improvement

{
  "dataset_size": 116,
  "elapsed_seconds": 3783.87,
  "epochs": 2,
  "export": {
    "adapter_dir": "/home/keith/gemma-v2/adapters/e4b_true_opd_v5",
    "keqv_v_proj_cloned_layers": 0,
    "safetensors_bytes": 279129832,
    "tensor_count": 516
  },
  "history_head": [
    {
      "active_tokens_cum": 21150,
      "ce_loss": 0.2605,
      "cispo_loss": 0.0521,
      "elapsed_sec": 477.7,
      "epoch": 1,
      "event": "opt_step",
      "grad_norm": 0.1543,
      "loss": 0.5034,
      "lr_a": 8e-07,
      "opt_step": 1,
      "rkl_loss": 0.6919,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 34886,
      "ce_loss": 0.386,
      "cispo_loss": 0.232,
      "elapsed_sec": 550.5,
      "epoch": 1,
      "event": "opt_step",
      "grad_norm": 0.2656,
      "loss": 0.7164,
      "lr_a": 1.6e-06,
      "opt_step": 2,
      "rkl_loss": 0.9375,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 53443,
      "ce_loss": 0.2444,
      "cispo_loss": 0.1966,
      "elapsed_sec": 626.1,
      "epoch": 1,
      "event": "opt_step",
      "grad_norm": 0.1279,
      "loss": 0.4566,
      "lr_a": 2.4e-06,
      "opt_step": 3,
      "rkl_loss": 0.5856,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    }
  ],
  "history_tail": [
    {
      "active_tokens_cum": 1057978,
      "ce_loss": 0.316,
      "cispo_loss": 0.1749,
      "elapsed_sec": 3575.9,
      "epoch": 2,
      "event": "opt_step",
      "grad_norm": 0.103,
      "loss": 0.3813,
      "lr_a": 4.8e-07,
      "opt_step": 54,
      "rkl_loss": 0.4154,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 1088127,
      "ce_loss": 0.269,
      "cispo_loss": 0.0,
      "elapsed_sec": 3649.5,
      "epoch": 2,
      "event": "opt_step",
      "grad_norm": 0.1094,
      "loss": 0.3199,
      "lr_a": 4.5e-07,
      "opt_step": 55,
      "rkl_loss": 0.3831,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 1101125,
      "ce_loss": 0.2748,
      "cispo_loss": 0.0511,
      "elapsed_sec": 3676.2,
      "epoch": 2,
      "event": "opt_step",
      "grad_norm": 0.0771,
      "loss": 0.3053,
      "lr_a": 4.3e-07,
      "opt_step": 56,
      "rkl_loss": 0.3445,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 1119460,
      "ce_loss": 0.2793,
      "cispo_loss": -0.255,
      "elapsed_sec": 3718.6,
      "epoch": 2,
      "event": "opt_step",
      "grad_norm": 0.0444,
      "loss": -0.0285,
      "lr_a": 4.1e-07,
      "opt_step": 57,
      "rkl_loss": 0.116,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    },
    {
      "active_tokens_cum": 1141764,
      "ce_loss": 0.4836,
      "cispo_loss": -0.4195,
      "elapsed_sec": 3783.7,
      "epoch": 2,
      "event": "opt_step",
      "grad_norm": 0.0608,
      "loss": -0.0562,
      "lr_a": 4e-07,
      "opt_step": 58,
      "rkl_loss": 0.1287,
      "total_opt_steps": 58,
      "unique_active_tokens": 570882
    }
  ],
  "injected_module_count": 258,
  "mode": "dual_gpu_opd",
  "opt_steps": 58,
  "student_alias": "gemma-4-e4b-it-qat-w4a16-ct",
  "total_active_token_updates": 1141764,
  "unique_active_tokens": 570882
}