penfever commited on
Commit
20a797c
·
verified ·
1 Parent(s): f5c98ae

Document current difficulty for every Good Atlas source

Browse files
Files changed (1) hide show
  1. README.md +1 -1
README.md CHANGED
@@ -44,7 +44,7 @@ The added tasks preserve upstream grading behavior and reference specific upstre
44
 
45
  MiMo's music and code configurations are included. **Code quality review is pending:** canonical no-op and empty-commit controls on one task matched separate direct native runs with reward 0; synthetic controls validate the adapter's behavior. This does not establish quality across all 2,698 tasks. MiMo's cyber, general and webdev configurations are not yet included; native runtime validation and publication remain pending.
46
 
47
- Difficulty coverage remains incomplete. As of 2026-10-06 at 20:03 UTC, the [public RL Data Atlas](https://public.applets.marina.oa.dev/a/fb11c931-5861-4878-8bb5-a964d652b45f/) has **current difficulty comparisons for 41 of its 42 sources rated Good** by its quality reviews. These source entries come from Task Trove and MarinSkyRL. Difficulty comparisons report each model's fraction of sampled tasks solved. Current comparisons use the same sampled tasks per source across Small Qwen3-Coder-30B-A3B-Instruct, Large Qwen3.5-122B-A10B (thinking disabled) and Hosted GLM-5.3 with low reasoning effort, graded by each source's current native grader. **Eurus code remains unmeasured:** the current native preparation rejects one of its 32 sampled records because its standard-input test data uses lists where strings are required. Native preparation succeeded for the other 31 records; the comparison remains unpublished while this failure is resolved. The preparation failure is retained and does not count as an incorrect model answer. Completing current difficulty runs for every Good source and publishing their results at the public endpoint remains part of the ongoing Atlas update's scope.
48
 
49
  ## Original Task Trove conversion
50
 
 
44
 
45
  MiMo's music and code configurations are included. **Code quality review is pending:** canonical no-op and empty-commit controls on one task matched separate direct native runs with reward 0; synthetic controls validate the adapter's behavior. This does not establish quality across all 2,698 tasks. MiMo's cyber, general and webdev configurations are not yet included; native runtime validation and publication remain pending.
46
 
47
+ As of 2026-10-06 at 21:05 UTC, the [public RL Data Atlas](https://public.applets.marina.oa.dev/a/fb11c931-5861-4878-8bb5-a964d652b45f/) has **current difficulty comparisons for all 42 sources rated Good** by its quality reviews. These source entries come from Task Trove and MarinSkyRL. Difficulty comparisons report each model's fraction of sampled tasks solved. Current comparisons use the same sampled tasks per source across Small Qwen3-Coder-30B-A3B-Instruct, Large Qwen3.5-122B-A10B (thinking disabled) and Hosted GLM-5.3 with low reasoning effort, graded by each source's current native grader. Completing and publishing these comparisons was part of the Atlas update's scope. **Eurus includes one unverified Small-model attempt:** its native sandbox lost state after its generated program reached the six-second execution time limit. The native error and model response are retained; the attempt receives no score. Eurus solve rates use 31 verified Small-model attempts and 32 each for Large and Hosted. Eurus's standard-input line lists were serialized with newlines using the source grader's documented behavior before loading tasks into the native environment; all 32 originally sampled records were retained.
48
 
49
  ## Original Task Trove conversion
50