aleleba 20493a99b6 Phase 6.4.25b: gate 1 on v2b-bf16 -- full PASS on all buckets ...

Global weighted loss 0.2725 (vs baseline 0.2750, previous v2-bf16 was
0.2734). All 6/6 non-penpot buckets pass the 0.10 regression threshold,
including delegacion_subagentes which had narrowly failed on v2-bf16
(+0.1019 -> now +0.0882). The 13 corrective seeds didn't hurt forgetting.

Also points vllm-eval at the new v2b-bf16 checkpoint for the gate 2/5
re-measurement.

2026-08-03 22:50:43 +00:00

..

gate5_png_prod-baseline

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_png_prod-baseline-r2

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_png_prod-baseline-r3

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_png_v2-bf16-full

Phase 6.4.25: gate 5 full-mode results on v2-bf16 (verdict: FAIL)

2026-08-03 13:33:15 +00:00

raw

Phase 6.4.25b: 13 corrective seeds for gate 5 diagnosis, rebuild training mix

2026-08-03 15:44:22 +00:00

schemas

Phase 6.4.25b: 13 corrective seeds for gate 5 diagnosis, rebuild training mix

2026-08-03 15:44:22 +00:00

calibration_v2.jsonl

Phase 6.4.25b: 13 corrective seeds for gate 5 diagnosis, rebuild training mix

2026-08-03 15:44:22 +00:00

chat_template_train.jinja

Fase 2: 06_validate_dataset.py - validacion con tokenizer/chat_template real (1470/1470 ok, 0 excepciones, 0 filtrados, 0 secretos)

2026-07-29 00:56:23 +00:00

eval_lora2.jsonl

Phase 6.4.25b: 13 corrective seeds for gate 5 diagnosis, rebuild training mix

2026-08-03 15:44:22 +00:00

eval.jsonl

Fase 2: 05_build_dataset.py - ensamblado v1 (1470 ejemplos: 750 nuevos + 720 replay, split 90/10 estratificado por bucket)

2026-07-29 00:50:10 +00:00

gate1_v2b-bf16_full.log

Phase 6.4.25b: gate 1 on v2b-bf16 -- full PASS on all buckets

2026-08-03 22:50:43 +00:00

gate2_results_bf16_2048.json

Fase 5: gate2 corregido (max_tokens=2048) - comparacion apples-to-apples

2026-07-30 07:32:50 +00:00

gate2_results_nvfp4_mixed_2048.json

Fase 5: gate2 corregido (max_tokens=2048) - comparacion apples-to-apples

2026-07-30 07:32:50 +00:00

gate2_results_nvfp4_mixed.json

Fase 5: puertas 2-3 sobre el checkpoint NVFP4 con calibracion mezclada

2026-07-30 06:20:58 +00:00

gate2_results_nvfp4_nospec.json

Fase 5: aislamiento - puertas 2-3 sin --speculative-config (mismo checkpoint)

2026-07-30 02:54:38 +00:00

gate2_results_nvfp4.json

Fase 5: resultados de las puertas 2-4 contra el checkpoint NVFP4+speculative

2026-07-30 00:45:50 +00:00

gate2_results_prod-baseline.json

Phase 6.3: re-measure the gate 2 and 3 production baselines on the new holdout

2026-07-30 21:25:58 +00:00

gate2_results_v2bf16.json

Phase 6.4.25: point vllm-eval at v2-bf16 and record gate 2/5-holdout results

2026-08-03 07:31:04 +00:00

gate2_results.json

Fase 4: resultados de las puertas 2, 3 y 4 sobre el checkpoint mergeado

2026-07-29 19:00:02 +00:00

gate3_results_nvfp4_mixed_2048.json

Fase 5: gate3 corregido (max_tokens=2048) - 100% (10/10), regresion era del test

2026-07-30 06:43:11 +00:00

gate3_results_nvfp4_mixed.json

Fase 5: puertas 2-3 sobre el checkpoint NVFP4 con calibracion mezclada

2026-07-30 06:20:58 +00:00

gate3_results_nvfp4_nospec.json

Fase 5: aislamiento - puertas 2-3 sin --speculative-config (mismo checkpoint)

2026-07-30 02:54:38 +00:00

gate3_results_nvfp4.json

Fase 5: resultados de las puertas 2-4 contra el checkpoint NVFP4+speculative

2026-07-30 00:45:50 +00:00

gate3_results_prod-baseline.json

Phase 6.3: re-measure the gate 2 and 3 production baselines on the new holdout

2026-07-30 21:25:58 +00:00

gate3_results.json

Fase 4: resultados de las puertas 2, 3 y 4 sobre el checkpoint mergeado

2026-07-29 19:00:02 +00:00

gate4_results_nvfp4.json

Fase 5: resultados de las puertas 2-4 contra el checkpoint NVFP4+speculative

2026-07-30 00:45:50 +00:00

gate4_results.json

Fase 4: resultados de las puertas 2, 3 y 4 sobre el checkpoint mergeado

2026-07-29 19:00:02 +00:00

gate5_holdout_v2-bf16-holdout.json

Phase 6.4.25: point vllm-eval at v2-bf16 and record gate 2/5-holdout results

2026-08-03 07:31:04 +00:00

gate5_paginas_prod-baseline-r2.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_paginas_prod-baseline-r3.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_paginas_prod-baseline.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_paginas_v2-bf16-full.json

Phase 6.4.25: gate 5 full-mode results on v2-bf16 (verdict: FAIL)

2026-08-03 13:33:15 +00:00

gate5_prod_baseline_run-SECCION-14pct.log

Phase 6.3.17: fix a harness fidelity bug and add a vibrancy metric

2026-07-30 18:38:24 +00:00

gate5_prod_baseline_run.log

Phase 6.3.17: fix a harness fidelity bug and add a vibrancy metric

2026-07-30 18:38:24 +00:00

gate5_prompts.jsonl

Phase 6.3.17: fix a harness fidelity bug and add a vibrancy metric

2026-07-30 18:38:24 +00:00

gate5_results_prod-baseline-AUDITROOT-BUG.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_results_prod-baseline-parcial-5de10.json

Phase 6.3.17: fix a harness fidelity bug and add a vibrancy metric

2026-07-30 18:38:24 +00:00

gate5_results_prod-baseline-r2.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_results_prod-baseline-r3.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_results_prod-baseline-SECCION-14pct.json

Phase 6.3.17: fix a harness fidelity bug and add a vibrancy metric

2026-07-30 18:38:24 +00:00

gate5_results_prod-baseline.json

Phase 6.3.17: close the production baseline with three repetitions

2026-07-30 21:01:59 +00:00

gate5_results_v2-bf16-full.json

Phase 6.4.25: gate 5 full-mode results on v2-bf16 (verdict: FAIL)

2026-08-03 13:33:15 +00:00

holdout_penpot_design.jsonl

Phase 6.3: fix the augmentation, close the holdout leak, stop rewarding invented parameters

2026-07-30 17:16:05 +00:00

holdout_prompts.jsonl

Phase 6.3: fix the augmentation, close the holdout leak, stop rewarding invented parameters

2026-07-30 17:16:05 +00:00

train_lora2_worstcase32.jsonl

Phase 6.4.20: build a worst-case probe set for the smoke run

2026-07-30 21:27:33 +00:00

train_lora2.jsonl

Phase 6.4.25b: 13 corrective seeds for gate 5 diagnosis, rebuild training mix

2026-08-03 15:44:22 +00:00

train.jsonl

Fase 2: 05_build_dataset.py - ensamblado v1 (1470 ejemplos: 750 nuevos + 720 replay, split 90/10 estratificado por bucket)

2026-07-29 00:50:10 +00:00