kobzaond/RLVRAMBench
RLVRAMBench Which language-model training configurations can I use with the memory I have, and how much testing does that decision require? RLVRAMBench is a measurement dataset with open evaluation tasks for a specific language-model training system. It measures memory feasibility when response generation and reinforcement-learning updates share the same graphics processors. It provides measured outcomes, fixed prediction tasks, a budgeted decision replay, reference methods, and… See the full description on the dataset page: https://huggingface.co/datasets/kobzaond/RLVRAMBench.
0223
1{2 "protocol_version": "1.0",3 "configurations": 212,4 "eligible_processes": 612,5 "recorded_attempts": 683,6 "never_started_conditions": 6,7 "stage_measurements": 3434,8 "historical_source_screens_counted_separately": 16,9 "transfer_configurations": 98,10 "distinct_target_configurations": 94,11 "queries": 400,12 "tasks": 52,13 "per_track": {14 "model_transfer": {15 "queries": 144,16 "tasks": 2417 },18 "workload_transfer": {19 "queries": 216,20 "tasks": 1221 },22 "gpu_count_transfer": {23 "queries": 36,24 "tasks": 1225 },26 "horizon_transfer": {27 "queries": 4,28 "tasks": 429 }30 },31 "eligible_processes_by_study": {32 "batch_logging_control": 24,33 "batch_reservation": 144,34 "boundary": 216,35 "four_gpu": 54,36 "horizon_followup": 24,37 "instrumentation": 72,38 "sampling_calibration": 12,39 "state_release": 36,40 "temporal_100": 12,41 "temporal_40": 1842 },43 "original_100_step_attempts": [44 {45 "attempt_id": "attempt-8e124d3b9f944cfd",46 "configuration_id": "cfg-331bb196cd92fe75",47 "run_id": "",48 "study": "temporal_100",49 "experiment_id": "strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92",50 "job_id": "5032255",51 "eligible": 0,52 "recorded_outcome": "no_trial_record",53 "documented_completion": 0,54 "failure_kind": "",55 "exclusion_reason": "temporal_cuda_ipc_weight_sync_failure",56 "last_completed_step_if_documented": 63,57 "evidence_files": "output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92/environment-5032255.json;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92/gpu-memory-5032255.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92/gpu-telemetry-5032255.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92/phase-memory-5032255.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-safe_margin-s92/training-5032255.log",58 "scheduler_evidence": ""59 },60 {61 "attempt_id": "attempt-1677fa3ecaffcc0b",62 "configuration_id": "cfg-c52a49b3491283c7",63 "run_id": "",64 "study": "temporal_100",65 "experiment_id": "strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91",66 "job_id": "5032257",67 "eligible": 0,68 "recorded_outcome": "no_trial_record",69 "documented_completion": 0,70 "failure_kind": "",71 "exclusion_reason": "temporal_weight_sync_no_progress",72 "last_completed_step_if_documented": 77,73 "evidence_files": "output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91/environment-5032257.json;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91/gpu-memory-5032257.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91/gpu-telemetry-5032257.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91/phase-memory-5032257.csv;output/strengthening_temporal_phases/strong-time-grpo-gsm8k-qwen25_3b-near_boundary-s91/training-5032257.log",74 "scheduler_evidence": ""75 }76 ]77}78