e4b_pure_opd_v7
When
Recorded date: 2026-11-03
First archive commit: 2026-11-04T17:35:03-05:00 (offset shown).
Training run day: 2026-11-03 (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?
Focus training examples by the model’s competence and add successful larger-model examples.
Which part changed?
Adapter matrices over the frozen base, targeting: ^model\.language_model\.layers\.\d+\.(self_attn\.(q_proj|k_proj|v_proj|o_proj)|mlp\.(gate_proj|up_proj|down_proj))$
The source describes adapter targets and mathematical changes. No before/after tensor comparison is available in this viewer.
How was it changed?
Competence-gated prediction distillation: 559 trajectories (245 student, 112 teacher, 202 unresolved); 2 epochs, 280 steps. Exported low-rank adapter, rank 64, scaling alpha 32.
Did it improve?
Historical measurements · Family results: epoch one 5/45, epoch two 6/45. The table does not certify which epoch this exact raw tensor export represents.
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
- v2/experiments/adapters/e4b_pure_opd_v7/checkpoint_manifest.json
- v2/experiments/adapters/e4b_pure_opd_v7/adapter_config.json
- docs/v2-post-training-and-lb-postmortem.md
- v2/experiments/adapters/e4b_pure_opd_v7/train_metrics.json
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 They may contain original predictions and superseded counts; use the measured-outcome summary and corrections above for the current result.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
| 9 | `e4b_pure_opd_v7` | Nov 3 | Competence-Gated (`+31B` pos) | `E4B` `r=64, a=32` | `opd_e4b_pure_v7` | `559` (`245` stu + `112` `31B` + `202` unres, `103` repos) | **`3.11M`** | `2` (`280`) | `6.23M` | `4.87h` | `ep1: 5/45`, `ep2: 6/45` | — | | 10 | `e4b_pure_opd_v7b` | Nov 3 | Pure On-Policy Competence-Gated | `E4B` `r=64, a=32` | `opd_e4b_pure_v7b` | `485` (`212` stu res + `273` unres, `101` repos) | `2.54M` | `2` (`243`) | `5.07M` | `2.99h` | `ep1_n200: 8/45`, `ep2_n200: 8/45` | — |
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_pure_opd_v7",
"wrx90_path": "/home/keith/gemma-v2/adapters/e4b_pure_opd_v7",
"files": {
"adapter_model.safetensors": {
"size_bytes": 279129832,
"sha256": "105fcd02808444f67496bac920ce06f83e23aa8ee9031366f3790d62e876068d"
}
},
"sub_checkpoints": {
"epoch_1": {
"adapter_model.safetensors": {
"size_bytes": 279129832
}
},
"epoch_2": {
"adapter_model.safetensors": {
"size_bytes": 279129832
}
}
}
}Training metrics — original archive, not task improvement
{
"dataset_size": 559,
"elapsed_seconds": 17518.55,
"epochs": 2,
"export": {
"adapter_dir": "/home/keith/gemma-v2/adapters/e4b_pure_opd_v7",
"keqv_v_proj_cloned_layers": 0,
"safetensors_bytes": 279129832,
"tensor_count": 516
},
"final_delta_w_frobenius_norm": 0.205,
"final_rms_layer_frobenius_norm": 0.012763,
"history_head": [
{
"active_tokens_cum": 30237,
"ce_loss": 0.4787,
"cispo_loss": 0.0,
"delta_w_frob": 0.00072,
"delta_w_pre_proj": 0.00072,
"delta_w_projected": false,
"elapsed_sec": 106.0,
"epoch": 1,
"event": "opt_step",
"grad_norm": 0.1994,
"loss": 0.5672,
"lr_a": 1.4e-07,
"opt_step": 1,
"rkl_loss": 0.7089,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 49287,
"ce_loss": 0.3909,
"cispo_loss": 0.0,
"delta_w_frob": 0.0017,
"delta_w_pre_proj": 0.0017,
"delta_w_projected": false,
"elapsed_sec": 157.0,
"epoch": 1,
"event": "opt_step",
"grad_norm": 0.2933,
"loss": 0.5785,
"lr_a": 2.9e-07,
"opt_step": 2,
"rkl_loss": 0.7231,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 71406,
"ce_loss": 0.3769,
"cispo_loss": 0.0,
"delta_w_frob": 0.0031,
"delta_w_pre_proj": 0.0031,
"delta_w_projected": false,
"elapsed_sec": 216.4,
"epoch": 1,
"event": "opt_step",
"grad_norm": 0.2169,
"loss": 0.6563,
"lr_a": 4.3e-07,
"opt_step": 3,
"rkl_loss": 0.8203,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
}
],
"history_tail": [
{
"active_tokens_cum": 6118025,
"ce_loss": 0.3376,
"cispo_loss": 0.0,
"delta_w_frob": 0.205,
"delta_w_pre_proj": 0.20622,
"delta_w_projected": true,
"elapsed_sec": 17280.7,
"epoch": 2,
"event": "opt_step",
"grad_norm": 0.4321,
"loss": 0.3626,
"lr_a": 4.1e-07,
"opt_step": 275,
"rkl_loss": 0.447,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 6154947,
"ce_loss": 0.4045,
"cispo_loss": 0.0,
"delta_w_frob": 0.205,
"delta_w_pre_proj": 0.20621,
"delta_w_projected": true,
"elapsed_sec": 17343.2,
"epoch": 2,
"event": "opt_step",
"grad_norm": 0.4455,
"loss": 0.4203,
"lr_a": 4e-07,
"opt_step": 276,
"rkl_loss": 0.5192,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 6183905,
"ce_loss": 0.438,
"cispo_loss": 0.0,
"delta_w_frob": 0.205,
"delta_w_pre_proj": 0.20621,
"delta_w_projected": true,
"elapsed_sec": 17406.5,
"epoch": 2,
"event": "opt_step",
"grad_norm": 0.4128,
"loss": 0.4216,
"lr_a": 4e-07,
"opt_step": 277,
"rkl_loss": 0.5212,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 6205502,
"ce_loss": 0.2251,
"cispo_loss": 0.0,
"delta_w_frob": 0.205,
"delta_w_pre_proj": 0.20621,
"delta_w_projected": true,
"elapsed_sec": 17466.8,
"epoch": 2,
"event": "opt_step",
"grad_norm": 0.5302,
"loss": 0.428,
"lr_a": 4e-07,
"opt_step": 278,
"rkl_loss": 0.5263,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
},
{
"active_tokens_cum": 6221244,
"ce_loss": 0.2518,
"cispo_loss": 0.0,
"delta_w_frob": 0.205,
"delta_w_pre_proj": 0.20623,
"delta_w_projected": true,
"elapsed_sec": 17505.8,
"epoch": 2,
"event": "opt_step",
"grad_norm": 0.5925,
"loss": 0.4345,
"lr_a": 4e-07,
"opt_step": 279,
"rkl_loss": 0.5347,
"total_opt_steps": 280,
"unique_active_tokens": 3114274
}
],
"injected_module_count": 258,
"max_qat_frobenius_norm": 0.205,
"mode": "dual_gpu_opd",
"opt_steps": 280,
"preflight_gate": {
"dataset_size": 559,
"eligible_unique_active_tokens": 3114274,
"epochs": 2,
"is_qat_model": true,
"kl_base_weight": 0.2,
"lora_scaling": 0.5,
"loraplus_ratio": 2.0,
"lr_a": 4e-06,
"lr_b": 8e-06,
"max_qat_frobenius_norm": 0.205,
"missing_superset_tasks": [],
"mode": "dual_gpu_opd",
"n_cispo_neg_rows": 0,
"n_resolved_rows": 357,
"n_unresolved_opd_rows": 202,
"opd_weight": 1.0,
"passed": true,
"planned_active_token_updates": 6228548,
"resolved_instances": 119,
"unique_instances": 128,
"violations": []
},
"student_alias": "gemma-4-e4b-it-qat-w4a16-ct",
"total_active_token_updates": 6228548,
"trust_region_projections": 248,
"unique_active_tokens": 3114274
}