Add SciCode Benchmark#1592
Merged
Merged
Conversation
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
…mark readme Signed-off-by: Frankie Siino <fsiino@nvidia.com>
- runner/rayexecutor + hdf5 target/compare helpers - app.py - config, schemas, verify() - fix configs - add requirements Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
- agent: run loop + step_utils (prompt context, code extraction, prefilled-steps data, context-window detection) + tests - prompts: default.yaml + background.yaml (agent-loaded) - scicode verify(): skips sub-steps absent from solutions Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
- SciCodeVerifyResponse carries problem_id into rollout output so rows are identifiable - agent config: example dataset - .gitignore adjustments - Add resources server example with metrics - Add benchmark metrics Signed-off-by: Frankie Siino <fsiino@nvidia.com>
- move agent instance + example dataset into resources server config - regenerate example_rollouts with example split Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
- resources_servers/scicode: replace scaffold - benchmarks/scicode: fix stale bits (metric, example path), add real-world benchmark run steps - scicode_agent: add prompt_fpath, note subtask_accuracy hook, fix prompt-default wording Signed-off-by: Frankie Siino <fsiino@nvidia.com>
gwarmstrong
requested changes
Jun 16, 2026
gwarmstrong
left a comment
gwarmstrong
left a comment
Contributor
There was a problem hiding this comment.
small request to add dataset license
| prepare_script: benchmarks/scicode/prepare.py | ||
| # null: the agent builds per-sub-step prompts (prompt_fpath); no row-level materialization. | ||
| prompt_config: null | ||
| num_repeats: 3 |
Contributor
There was a problem hiding this comment.
Can you add the dataset's license here?
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
Signed-off-by: Frankie Siino <fsiino@nvidia.com>
gwarmstrong
approved these changes
Jun 16, 2026
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.
Migrates the SciCode benchmark from nemo-skills into nemo-gym.
How it works in gym:
Architecture - Multi-step: the agent makes one model call per sub-step, accumulating its own prior-step code as context, then submits the accumulated solutions to
/verify. N model calls per problem (N varies).Execution - Resources server runs each sub-step's accumulated code in a Ray subprocess against ground-truth targets in
test_data.h5.Test data -
test_data.h5must be staged manually from the official SciCode Google Drive.Reward - Binary per problem (
1.0iff every sub-step passes), same as skills' problem-level accuracy.Headline metric -
subtask_accuracy(total sub-steps passed / total) viacompute_metrics/get_key_metricsoverride on the agent. Numerically equal to nemo-skills'pass@1[avg-of-3]/subtask_accuracy.Test results:
With ultra mopd/step36 checkpoint:
wandb: https://wandb.ai/nvidia/fsiino-gym-dev/runs/3jom33yh