Trained / exported adapter / adapter-m31b_rft_r32_v1

m31b_rft_r32_v1

gemma-4-31b-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?

Fine-tune the larger model on its own successful repair examples.

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?

31B self-repair fine-tuning: 26 solved trajectories, 2 epochs, 13 optimizer steps; 820 exported tensors. Exported low-rank adapter, rank 32, scaling alpha 64.

Did it improve?

Historical measurements · Historical local result: 8–9/15 on val15 and 3/8 on cross-validation.

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 They may contain original predictions and superseded counts; use the measured-outcome summary and corrections above for the current result.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

| 3 | `m31b_rft_r32_v1` | Nov 1 | `31B` Self-RFT (`820` tensors) | `31B` `r=32, a=16` | `rft_v2` | `26` (`26` `31B` solved, `11` tasks) | `0.24M` | `2` (`13`) | `0.47M` | `1.59h` | `8–9/15` (`val15`), `3/8` CV | — |

Adapter configuration — original metadata

{
  "base_model_name_or_path": "gemma-4-31b-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": "m31b_rft_r32_v1",
  "wrx90_path": "/home/keith/gemma-v2/adapters/m31b_rft_r32_v1",
  "files": {
    "adapter_model.safetensors": {
      "size_bytes": 494596016,
      "sha256": "423a07fb3c232fc163f981184c6d2668f660a23befd3dc9df3ca891d67a8d55e"
    }
  }
}

Training metrics — original archive, not task improvement

{
  "dataset_size": 26,
  "elapsed_seconds": 5738.95,
  "epochs": 2,
  "export": {
    "adapter_dir": "/home/keith/gemma-v2/adapters/m31b_rft_r32_v1",
    "safetensors_bytes": 489841600,
    "tensor_count": 820
  },
  "history_head": [
    {
      "ce_loss": 0.1216,
      "epoch": 1,
      "grad_norm": 1.1406,
      "loss": 0.1216,
      "lr_a": 1e-05,
      "opt_step": 1,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.1357,
      "epoch": 1,
      "grad_norm": 0.9141,
      "loss": 0.1357,
      "lr_a": 1e-05,
      "opt_step": 2,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.1881,
      "epoch": 1,
      "grad_norm": 0.8164,
      "loss": 0.1881,
      "lr_a": 9.87e-06,
      "opt_step": 3,
      "rkl_loss": 0.0
    }
  ],
  "history_tail": [
    {
      "ce_loss": 0.1688,
      "epoch": 2,
      "grad_norm": 0.2734,
      "loss": 0.1688,
      "lr_a": 4.96e-06,
      "opt_step": 9,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.1528,
      "epoch": 2,
      "grad_norm": 0.5234,
      "loss": 0.1528,
      "lr_a": 3.9e-06,
      "opt_step": 10,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.0925,
      "epoch": 2,
      "grad_norm": 0.5,
      "loss": 0.0925,
      "lr_a": 2.94e-06,
      "opt_step": 11,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.129,
      "epoch": 2,
      "grad_norm": 3.5,
      "loss": 0.129,
      "lr_a": 2.13e-06,
      "opt_step": 12,
      "rkl_loss": 0.0
    },
    {
      "ce_loss": 0.1656,
      "epoch": 2,
      "grad_norm": 0.9414,
      "loss": 0.1656,
      "lr_a": 1.52e-06,
      "opt_step": 13,
      "rkl_loss": 0.0
    }
  ],
  "injected_module_count": 410,
  "mode": "rft_lora",
  "opt_steps": 13,
  "student_alias": "gemma-4-31b-it-qat-w4a16-ct"
}