| { |
| "schema_version": 1, |
| "bootstrap_samples": 20000, |
| "bootstrap_seed": 20260817, |
| "protocol": { |
| "temperature": 0.6, |
| "top_p": 0.95, |
| "top_k": 20, |
| "maximum_agent_turns": 20, |
| "seed_base": 20260811, |
| "per_question_seed": "seed_base + benchmark index", |
| "generation_concurrency_per_gpu": 2, |
| "judge_model": "gemini-3.1-pro-preview", |
| "primary_denominator": "all 256 questions" |
| }, |
| "runs": { |
| "v1.0.2_historical_base": { |
| "results": "historical/v1.0.2/qwen3-8b/final_results.jsonl", |
| "summary": "historical/v1.0.2/qwen3-8b/summary.json", |
| "model_label": "Qwen/Qwen3-8B", |
| "database_revision": "v1.0.2", |
| "database_sha256": "7e5cd26ca7be16be1ff41962391478122624e48e5c6b3b91fb18fcccfbe2171e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2215_[0-3]", |
| "classification": "published historical reference from v1.0.2" |
| }, |
| "v1.0.2_historical_rl_iter199": { |
| "results": "historical/v1.0.2/browsecomp-rl-iter199/final_results.jsonl", |
| "summary": "historical/v1.0.2/browsecomp-rl-iter199/summary.json", |
| "model_label": "Shangy/browsecomp-qwen3-8B-length-penalty-iter199", |
| "database_revision": "v1.0.2", |
| "database_sha256": "7e5cd26ca7be16be1ff41962391478122624e48e5c6b3b91fb18fcccfbe2171e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2219_[0-3]", |
| "classification": "published historical reference from v1.0.2" |
| }, |
| "v1.0.2_fresh_base": { |
| "results": "controls/v1.0.2-fresh/qwen3-8b/final_results.jsonl", |
| "summary": "controls/v1.0.2-fresh/qwen3-8b/summary.json", |
| "model_label": "Qwen/Qwen3-8B", |
| "database_revision": "v1.0.2", |
| "database_sha256": "7e5cd26ca7be16be1ff41962391478122624e48e5c6b3b91fb18fcccfbe2171e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2289_[0-3]", |
| "classification": "fresh old-database control for the robots repair" |
| }, |
| "v1.0.2_fresh_rl_iter199": { |
| "results": "controls/v1.0.2-fresh/browsecomp-rl-iter199/final_results.jsonl", |
| "summary": "controls/v1.0.2-fresh/browsecomp-rl-iter199/summary.json", |
| "model_label": "Shangy/browsecomp-qwen3-8B-length-penalty-iter199", |
| "database_revision": "v1.0.2", |
| "database_sha256": "7e5cd26ca7be16be1ff41962391478122624e48e5c6b3b91fb18fcccfbe2171e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2290_[0-3]", |
| "classification": "fresh old-database control for the robots repair" |
| }, |
| "v1.1.0_replica_8x32_base": { |
| "results": "replicas/v1.1.0-8x32/qwen3-8b/final_results.jsonl", |
| "summary": "replicas/v1.1.0-8x32/qwen3-8b/summary.json", |
| "model_label": "Qwen/Qwen3-8B", |
| "database_revision": "v1.1.0", |
| "database_sha256": "09172388ed6c0ae91cb43dff3a663bcfb219673ed991ff5c9922904eb61ee77e", |
| "topology": "8 shards x 32 questions", |
| "slurm_array": "2379_[0-7]", |
| "classification": "complete topology/scheduling replica; non-canonical" |
| }, |
| "v1.1.0_replica_8x32_rl_iter199": { |
| "results": "replicas/v1.1.0-8x32/browsecomp-rl-iter199/final_results.jsonl", |
| "summary": "replicas/v1.1.0-8x32/browsecomp-rl-iter199/summary.json", |
| "model_label": "Shangy/browsecomp-qwen3-8B-length-penalty-iter199", |
| "database_revision": "v1.1.0", |
| "database_sha256": "09172388ed6c0ae91cb43dff3a663bcfb219673ed991ff5c9922904eb61ee77e", |
| "topology": "8 shards x 32 questions", |
| "slurm_array": "2380_[0-7]", |
| "classification": "complete topology/scheduling replica; non-canonical" |
| }, |
| "v1.1.0_matched_base": { |
| "results": "../../baselines/qwen3-8b/final_results.jsonl", |
| "summary": "../../baselines/qwen3-8b/summary.json", |
| "model_label": "Qwen/Qwen3-8B", |
| "database_revision": "v1.1.0", |
| "database_sha256": "09172388ed6c0ae91cb43dff3a663bcfb219673ed991ff5c9922904eb61ee77e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2395_[0-3]", |
| "classification": "canonical v1.1.0 matched baseline" |
| }, |
| "v1.1.0_matched_rl_iter199": { |
| "results": "../../baselines/browsecomp-rl-iter199/final_results.jsonl", |
| "summary": "../../baselines/browsecomp-rl-iter199/summary.json", |
| "model_label": "Shangy/browsecomp-qwen3-8B-length-penalty-iter199", |
| "database_revision": "v1.1.0", |
| "database_sha256": "09172388ed6c0ae91cb43dff3a663bcfb219673ed991ff5c9922904eb61ee77e", |
| "topology": "4 shards x 64 questions", |
| "slurm_array": "2396_[0-3]", |
| "classification": "canonical v1.1.0 matched baseline" |
| } |
| }, |
| "comparisons": [ |
| { |
| "name": "v1.0.2_repeatability_base", |
| "left": "v1.0.2_historical_base", |
| "right": "v1.0.2_fresh_base", |
| "interpretation": "same model, database, protocol and nominal topology; estimates run/serving variance" |
| }, |
| { |
| "name": "v1.0.2_repeatability_rl_iter199", |
| "left": "v1.0.2_historical_rl_iter199", |
| "right": "v1.0.2_fresh_rl_iter199", |
| "interpretation": "same model, database, protocol and nominal topology; estimates run/serving variance" |
| }, |
| { |
| "name": "robots_repair_matched_base", |
| "left": "v1.0.2_fresh_base", |
| "right": "v1.1.0_matched_base", |
| "interpretation": "primary matched corpus-repair comparison for the base model" |
| }, |
| { |
| "name": "robots_repair_matched_rl_iter199", |
| "left": "v1.0.2_fresh_rl_iter199", |
| "right": "v1.1.0_matched_rl_iter199", |
| "interpretation": "primary matched corpus-repair comparison for the RL checkpoint" |
| }, |
| { |
| "name": "v1.0.2_fresh_rl_gain", |
| "left": "v1.0.2_fresh_base", |
| "right": "v1.0.2_fresh_rl_iter199", |
| "interpretation": "RL versus base on the fresh old-database controls" |
| }, |
| { |
| "name": "v1.1.0_canonical_rl_gain", |
| "left": "v1.1.0_matched_base", |
| "right": "v1.1.0_matched_rl_iter199", |
| "interpretation": "RL versus base on the repaired corpus; primary model comparison" |
| }, |
| { |
| "name": "v1.1.0_replica_to_matched_base", |
| "left": "v1.1.0_replica_8x32_base", |
| "right": "v1.1.0_matched_base", |
| "interpretation": "same corpus/model/protocol, different shard topology and serving schedule" |
| }, |
| { |
| "name": "v1.1.0_replica_to_matched_rl_iter199", |
| "left": "v1.1.0_replica_8x32_rl_iter199", |
| "right": "v1.1.0_matched_rl_iter199", |
| "interpretation": "same corpus/model/protocol, different shard topology and serving schedule" |
| } |
| ], |
| "retrieval_stratification": { |
| "source": "../original_question_top20_comparison.json", |
| "comparisons": [ |
| { |
| "name": "robots_repair_matched_base", |
| "left": "v1.0.2_fresh_base", |
| "right": "v1.1.0_matched_base" |
| }, |
| { |
| "name": "robots_repair_matched_rl_iter199", |
| "left": "v1.0.2_fresh_rl_iter199", |
| "right": "v1.1.0_matched_rl_iter199" |
| } |
| ] |
| } |
| } |
|
|
