Trained / exported adapter / adapter-e4b_kl_r32_v2

e4b_kl_r32_v2

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?

Train on successful student repairs while penalizing departure from the base model.

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?

Student repair fine-tuning with a base-model prediction penalty weighted 0.2; 14 trajectories, 2 epochs, 7 optimizer steps. Exported low-rank adapter, rank 32, scaling alpha 16.

Did it improve?

Historical measurements · Historical local result: 6–7/15 on val15 and 1–2/8 on the cross-validation set.

Identical tensor-file bytes also appear under: e4b_opd_r32_v1. These are not separate demonstrated weight changes.

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 They may contain original predictions and superseded counts; use the measured-outcome summary and corrections above for the current result.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

| 2 | `e4b_kl_r32_v2` | Nov 1 | Student RFT + `0.2` base KL | `E4B` `r=32, a=16` | `rft_e4b_aligned_v2` | `14` (`14` `E4B` solved, `8` tasks) | `0.12M` | `2` (`7`) | `0.24M` | `0.17h` | `6–7/15` (`val15`), `1–2/8` CV | — |

Adapter configuration — original metadata

{
  "base_model_name_or_path": "gemma-4-e4b-it-qat-w4a16-ct",
  "bias": "none",
  "lora_alpha": 16,
  "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_kl_r32_v2",
  "wrx90_path": "/home/keith/gemma-v2/adapters/e4b_kl_r32_v2",
  "files": {
    "adapter_model.safetensors": {
      "size_bytes": 139603296,
      "sha256": "89ee3898c54bcd962d9aef7c2c7d050b2e091ff674f2959fcacdc05642c19ff4"
    }
  }
}

Training metrics — original archive, not task improvement

{
  "dataset_size": 14,
  "elapsed_seconds": 609.21,
  "epochs": 2,
  "export": {
    "adapter_dir": "/home/keith/gemma-v2/adapters/e4b_kl_r32_v2",
    "keqv_v_proj_cloned_layers": 0,
    "safetensors_bytes": 139603296,
    "tensor_count": 516
  },
  "history_head": [
    {
      "ce_loss": 0.4037,
      "epoch": 1,
      "grad_norm": 0.0752,
      "loss": 0.2018,
      "lr_a": 4e-06,
      "opt_step": 1,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.4836,
      "epoch": 1,
      "grad_norm": 0.0737,
      "loss": 0.2419,
      "lr_a": 4e-06,
      "opt_step": 2,
      "rkl_loss": 0.0002
    },
    {
      "ce_loss": 0.3609,
      "epoch": 1,
      "grad_norm": 0.0718,
      "loss": 0.1805,
      "lr_a": 3.82e-06,
      "opt_step": 3,
      "rkl_loss": 0.0001
    }
  ],
  "history_tail": [
    {
      "ce_loss": 0.3609,
      "epoch": 1,
      "grad_norm": 0.0718,
      "loss": 0.1805,
      "lr_a": 3.82e-06,
      "opt_step": 3,
      "rkl_loss": 0.0001
    },
    {
      "ce_loss": 0.5228,
      "epoch": 2,
      "grad_norm": 0.0786,
      "loss": 0.2615,
      "lr_a": 3.32e-06,
      "opt_step": 4,
      "rkl_loss": 0.0002
    },
    {
      "ce_loss": 0.3609,
      "epoch": 2,
      "grad_norm": 0.0713,
      "loss": 0.1805,
      "lr_a": 2.6e-06,
      "opt_step": 5,
      "rkl_loss": 0.0001
    },
    {
      "ce_loss": 0.3354,
      "epoch": 2,
      "grad_norm": 0.0635,
      "loss": 0.1678,
      "lr_a": 1.8e-06,
      "opt_step": 6,
      "rkl_loss": 0.0001
    },
    {
      "ce_loss": 0.4898,
      "epoch": 2,
      "grad_norm": 0.0757,
      "loss": 0.245,
      "lr_a": 1.08e-06,
      "opt_step": 7,
      "rkl_loss": 0.0002
    }
  ],
  "injected_module_count": 258,
  "mode": "rft_lora",
  "opt_steps": 7,
  "student_alias": "gemma-4-e4b-it-qat-w4a16-ct"
}