YWZBrandon's picture
Checkpoint evaluation plot data, snapshot 2026-09-14
2ba23ae verified
Raw History Blame Contribute Delete
10.7 kB
reference_id,figure,dataset,role,label,checkpoint_iteration,score_mean,steps_mean,output_tokens_mean,step_cap,attempts_per_task,source_file,ci_status,confidence_intervals,precision,source_kind,notes,displayed_at_export,temperature,tasks,episodes,search_file_policy,source_result,reference_snapshot_source,reference_captured_utc
opd_hill_climb/reference-0,opd_hill_climb,officeqa,horizontal_reference,397B teacher,,0.5093984962406015,36.025375939849624,19922.35620300752,120,,notes/figures/opd_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
opd_hill_climb/reference-1,opd_hill_climb,officeqa,horizontal_reference,SFT all-data iter-69 (cap 120),,0.301,57.2,30700.0,120,,notes/figures/opd_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
opd_hill_climb/reference-2,opd_hill_climb,officeqa,horizontal_reference,Qwen3.5-4B base (cap 120),,0.132,55.9,24400.0,120,,notes/figures/opd_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
opd_hill_climb/step-0,opd_hill_climb,officeqa,shared_step_zero,Qwen3.5-4B base,0,0.13157894736842105,55.90695488721804,24428.878759398496,120,8,notes/figures/opd_hill_climb.py,available,"{""output_tokens"": {""low"": 22596.956766917294, ""high"": 26297.461466165412, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}, ""score"": {""low"": 0.11258607228492401, ""high"": 0.15322263941913394, ""level"": 0.95, ""method"": ""Wilson, z=1.96, over attempts (not task-cluster adjusted)""}, ""steps"": {""low"": 52.18796992481203, ""high"": 59.58646616541353, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}}",as written in plot source,offline_base_reference,"One shared baseline, not an independent evaluation for every training arm.",,,,,,,,
q38_hill_climb/reference-0,q38_hill_climb,officeqa,horizontal_reference,Qwen3.8-27B teacher,,0.648,15.6,17600.0,120,,notes/figures/q38_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
q38_hill_climb/reference-1,q38_hill_climb,officeqa,horizontal_reference,Qwen3.5-4B base (cap 120),,0.132,55.9,24400.0,120,,notes/figures/q38_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
q38_hill_climb/step-0,q38_hill_climb,officeqa,shared_step_zero,Qwen3.5-4B base,0,0.13157894736842105,55.90695488721804,24428.878759398496,120,8,notes/figures/q38_hill_climb.py,available,"{""output_tokens"": {""low"": 22596.956766917294, ""high"": 26297.461466165412, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}, ""score"": {""low"": 0.11258607228492401, ""high"": 0.15322263941913394, ""level"": 0.95, ""method"": ""Wilson, z=1.96, over attempts (not task-cluster adjusted)""}, ""steps"": {""low"": 52.18796992481203, ""high"": 59.58646616541353, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}}",as written in plot source,offline_base_reference,"One shared baseline, not an independent evaluation for every training arm.",,,,,,,,
label_hill_climb/reference-0,label_hill_climb,officeqa,horizontal_reference,Qwen3.5-397B teacher (cap 120),,0.5093984962406015,36.025375939849624,19922.35620300752,120,,notes/figures/label_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
label_hill_climb/reference-1,label_hill_climb,officeqa,horizontal_reference,Qwen3.5-4B base (cap 120),,0.13157894736842105,55.90695488721804,24428.878759398496,120,,notes/figures/label_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
label_hill_climb/step-0,label_hill_climb,officeqa,shared_step_zero,Qwen3.5-4B base,0,0.13157894736842105,55.90695488721804,24428.878759398496,120,8,notes/figures/label_hill_climb.py,not_in_reference_source,,as written in plot source,offline_base_reference,"One shared baseline, not an independent evaluation for every training arm.",,,,,,,,
ssb1_hill_climb/reference-0,ssb1_hill_climb,ssb1,horizontal_reference,Qwen3.5-4B base,,0.318,26.5,14300.0,50,1,notes/figures/ssb1_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,"Legacy pass1 base reference, despite the sweep subtitle saying pass@8.",,,,,,,,
ssb1_hill_climb/reference-1,ssb1_hill_climb,ssb1,horizontal_reference,397B teacher,,0.773125,12.5403125,8621.975,50,,notes/figures/ssb1_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
ssb1_hill_climb/reference-2,ssb1_hill_climb,ssb1,horizontal_reference,SFT all-data iter-309,,0.4978125,23.7896875,24752.4578125,50,,notes/figures/ssb1_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
ssb1_hill_climb/step-0,ssb1_hill_climb,ssb1,shared_step_zero,Qwen3.5-4B base,0,0.318,26.5,14300.0,50,1,notes/figures/ssb1_hill_climb.py,not_in_reference_source,,as written in plot source,offline_base_reference,"One shared baseline, not an independent evaluation for every training arm.",,,,,,,,
ssb1_label_hill_climb/reference-0,ssb1_label_hill_climb,ssb1,horizontal_reference,Qwen3.5-4B base,,0.318,26.5,14300.0,50,1,notes/figures/ssb1_label_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,"Legacy pass1 base reference, despite the sweep subtitle saying pass@8.",,,,,,,,
ssb1_label_hill_climb/step-0,ssb1_label_hill_climb,ssb1,shared_step_zero,Qwen3.5-4B base,0,0.318,26.5,14300.0,50,1,notes/figures/ssb1_label_hill_climb.py,not_in_reference_source,,as written in plot source,offline_base_reference,"One shared baseline, not an independent evaluation for every training arm.",,,,,,,,
gemma_hill_climb/reference-0,gemma_hill_climb,officeqa,horizontal_reference,Gemma-4-31B teacher (cap 120),,0.4023,22.8,31270.0,120,,notes/figures/gemma_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
gemma_hill_climb/reference-1,gemma_hill_climb,officeqa,horizontal_reference,Gemma-4-12B base (cap 120),,0.1363,23.8,32870.0,120,,notes/figures/gemma_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
jobbench_mtrain_hill_climb/reference-0,jobbench_mtrain_hill_climb,jobbench_easy,horizontal_reference,"397B teacher (cap 50, temp 1.0)",,0.5331,21.6,,50,,notes/figures/jobbench_mtrain_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,,,,,,,,,
jobbench_mtrain_hill_climb/reference-1,jobbench_mtrain_hill_climb,jobbench_easy,conditional_fallback_reference,"Qwen3.5-4B base (cap 50, temp 1.0)",,0.2214,36.4,,50,,notes/figures/jobbench_mtrain_hill_climb.py,not_in_reference_source,,as written in plot source (may be rounded),offline_reference,Only displayed if the validated cap-120 base is unavailable.,False,,,,,,,
qwen9b_hill_climb/teacher,qwen9b_hill_climb,officeqa,horizontal_reference,397B teacher (cap 120),,0.5093984962406015,36.025375939849624,19922.35620300752,120,8,notes/figures/officeqa_teacher_reference.py,not_in_reference_source,,,offline_reference,,,,,,,,,
ssb1_qwen9b_hill_climb/teacher,ssb1_qwen9b_hill_climb,ssb1,horizontal_reference,397B teacher,,0.773,12.54,8620.0,50,8,notes/figures/ssb1_qwen9b_hill_climb.py,not_in_reference_source,,,offline_reference,,,,,,,,,
qwen27b_hill_climb/teacher,qwen27b_hill_climb,officeqa,horizontal_reference,397B teacher (cap 120),,0.5093984962406015,36.025375939849624,19922.35620300752,120,8,notes/figures/officeqa_teacher_reference.py,not_in_reference_source,,,offline_reference,,,,,,,,,
ssb1_qwen27b_hill_climb/teacher,ssb1_qwen27b_hill_climb,ssb1,horizontal_reference,397B teacher,,0.773,12.54,8620.0,50,8,notes/figures/ssb1_qwen27b_hill_climb.py,not_in_reference_source,,,offline_reference,,,,,,,,,
jobbench_mtrain_hill_climb/step-0,jobbench_mtrain_hill_climb,jobbench_easy,shared_step_zero_and_horizontal_reference,"Qwen3.5-4B base (cap 120, staged)",0,0.2845843253968254,41.660714285714285,30173.880952380943,120,8,notes/figures/jobbench_base_reference.py,available,"{""score"": {""low"": 0.23949801587301583, ""high"": 0.33214880952380954, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}, ""output_tokens"": {""low"": 26210.230158730155, ""high"": 34542.04365079365, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}, ""steps"": {""low"": 36.964285714285715, ""high"": 46.70436507936508, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}}",,offline_base_reference,Validated all 504 saved attempts and staged protocol using the plot's existing read-only baseline loader; no raw attempt content included. The conditional cap-50 base fallback is not displayed.,,1.0,63,504,stage,shared-bucket/officeqa-rl/eval-results/jobbench-easy-vllm-Qwen3.5-4B-base-pass8-step120-temp1-sfstage-dp8-tp1,results/checkpoint_sweeps/plot_base_reference_snapshot_20260914.json,2026-09-14T16:23:57+00:00
qwen9b_hill_climb/step-0,qwen9b_hill_climb,officeqa,shared_step_zero_and_horizontal_reference,Qwen3.5-9B base (cap 120),0,0.17481203007518797,42.964285714285715,27006.141917293233,120,8,notes/figures/qwen9b_base_reference.py,available,"{""score"": {""low"": 0.15317130788522437, ""high"": 0.19879250397601045, ""level"": 0.95, ""method"": ""Wilson, z=1.96, over attempts (not task-cluster adjusted)""}, ""output_tokens"": {""low"": 24725.5977443609, ""high"": 29280.967105263157, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}, ""steps"": {""low"": 39.87781954887218, ""high"": 45.971804511278194, ""level"": 0.95, ""method"": ""task-cluster percentile bootstrap, 10000 draws, 95%""}}",,offline_base_reference,"Completed summary: 186 correct / 1064 attempts, 28734535 output tokens, 45714 calls. Token/step intervals are cached q9bbase-iter0-pass8-cap120 intervals.",,1.0,133,1064,,shared-bucket/officeqa-rl/eval-results/officeqa-pro-vllm-Qwen3.5-9B-base-pass8-step120-temp1-dp8-tp1-text-docker,results/checkpoint_sweeps/plot_base_reference_snapshot_20260914.json,2026-09-14T16:23:57+00:00
qwen27b_hill_climb/legacy-base,qwen27b_hill_climb,officeqa,horizontal_reference,"27B base (in-training, cap 50)",,0.3458646616541353,28.32330827067669,11188.66917293233,50,,notes/figures/series/series_q27b_pure.tsv,not_in_reference_source,,,in_training_reference,Plot uses raw len/steps training columns; not a controlled offline step-0 evaluation.,,,,,,,,