Repository navigation
test: update reward model expected score - #4252
Merged
Merged
Conversation
Signed-off-by: Yuki Huang <yukih@nvidia.com>
Contributor
Author
|
/ok to test 31a5dc5 |
terrykong
approved these changes
Sep 23, 2026
terrykong
enabled auto-merge (squash)
September 23, 2026 17:00
cspades
pushed a commit
that referenced
this pull request
Sep 24, 2026
Signed-off-by: Yuki Huang <yukih@nvidia.com>
yfw
added a commit
that referenced
this pull request
Sep 24, 2026
Brings #4052 (single-controller telemetry; moves nemo-lens to b0f977d and adds the aiohttp extra), #4252 (reward-model environment expected score, which fixes `L0_Unit_Tests_Environments` on the previous merge), #4248, #4234, #4254 and #4171. Only uv.lock conflicted: regenerated from this branch's lock; nemo-lens moves to main's rev and the three opentelemetry-instrumentation packages its aiohttp extra needs are added. Nothing else moved, and nothing resolves below main except the deliberate numpy / llguidance / flash-attn pins. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Yi-Fu Wu <yifu.wu@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Update the expected incorrect-answer reward in
test_reward_model_environment_generate_rewardsfrom -5.2500 to -5.4062.Since #4143, frozen forward-only models (
init_optimizer=False, incl. the reward model) load directly in the runtime dtype (bf16) instead of fp32. The Skywork RM checkpoint is all-BF16, so the weights used in compute are identical. Only the incorrect-answer score moves (by 0.156); the correct-answer score (2.6719) and the correct > incorrect ordering are unchanged.-5.4062 is not a new value. The existing test comment already lists it as one of the observed scores (-5.2500 / -5.3750 / -5.4062 across environments). This PR only switches which of those observed values CI asserts against.
Main has failed deterministically with -5.4062 since cba6952, the first CI run after #4143 merged.