| { |
| "combined_peak_lr": 1.5e-05, |
| "development_cases": 300, |
| "fifth_trial": "Choose original starting checkpoint using matched repair trials 1/2 dev scores; allocate 10%-40% per domain using positive dev gains from first four. No chaining of experimental checkpoints.", |
| "final_lr_fraction": 0.1, |
| "legacy_dev_cases": 320, |
| "legacy_test_cases": 704, |
| "maximum_attempts": 5, |
| "new_objective_weight": 0.4, |
| "peak_lr": 2e-05, |
| "practical_success": "At least 5pp new heldout content correctness gain over EACH baseline with 97.5% paired scenario-bootstrap intervals above zero; preserve strict instruction compliance, domains and general benchmarks within predeclared tolerances.", |
| "preservation": "Both starting models and previous protected checkpoints remain immutable. Keep the highest-dev-ranked new candidate even if final results do not establish an improvement. Remove other new trial artifacts only after all final assessments complete.", |
| "public_tolerances": { |
| "chat_answer_cases": 2, |
| "chat_repetition_cases": 1, |
| "chat_stop_cases": 1, |
| "choice_mean_pp": 0.25, |
| "choice_per_task_pp": 1, |
| "ifeval_loose_pp": 0.5, |
| "ifeval_strict_pp": 0 |
| }, |
| "replay_kl_weight": 0.25, |
| "replay_objective_weight": 0.6, |
| "reserved_test_cases": 1000, |
| "seed": 2026110915, |
| "selected_new_conversations": 4096, |
| "selected_replay_conversations": 4096, |
| "selection": "Highest dev key with preservation guards; one finalist is frozen before reserved test or new public benchmark evaluation.", |
| "trials": [ |
| { |
| "mixture": { |
| "abstention": 0.4, |
| "dialogue": 0.2, |
| "grounded": 0.4 |
| }, |
| "name": "Extraction and appropriate abstention", |
| "parent": "attempt5" |
| }, |
| { |
| "mixture": { |
| "abstention": 0.4, |
| "dialogue": 0.2, |
| "grounded": 0.4 |
| }, |
| "name": "Matched repair from parent2000", |
| "parent": "parent2000" |
| }, |
| { |
| "mixture": { |
| "abstention": 0.1, |
| "constraints": 0.65, |
| "dialogue": 0.15, |
| "grounded": 0.1 |
| }, |
| "name": "Practical combined instructions", |
| "parent": "attempt5" |
| }, |
| { |
| "mixture": { |
| "abstention": 0.1, |
| "dialogue": 0.15, |
| "grounded": 0.15, |
| "reasoning": 0.6 |
| }, |
| "name": "Short checked reasoning", |
| "parent": "attempt5" |
| }, |
| { |
| "mixture": null, |
| "name": "Development-selected combination", |
| "parent": null |
| } |
| ], |
| "updates_per_attempt": 256, |
| "warmup_fraction": 0.1 |
| } |
|
|
