Closed
Conversation
Implements a Nemo-Gym resources server for verifiable chemistry question answering (direct generation variant). The agent receives a natural-language chemistry question + SMILES and must respond with a single number or binary flag. The reward signal mirrors the offline benchmarking pipeline: exact-match for integer/bool/fragment properties; percentile-based threshold (must beat 95% of the ChEMBL prior) for continuous float properties. Data generated from the full_strat025_perm benchmark experiment via scripts/export_nemo_gym_data.py and scripts/precompute_reward_stats.py in chemistry-benchmarking-for-lead-op (branch: dane/nemo-gym-rl-port). Made-with: Cursor
Replace the percentile-based binary (0/1) float reward with a continuous negative absolute error: reward = -|predicted - actual|. A perfect prediction scores 0.0; larger errors give more negative rewards. This removes the dependency on reward_stats.json and features.parquet quantile precomputation. - app.py: remove quantile machinery (numpy, FLOAT_ACCURACY_THRESHOLD, _float_reward, _load_reward_stats, ChemistryDirectConfig.reward_stats_path); simplify compute_reward() to three-line dispatch; add absolute_error field to ChemistryDirectVerifyResponse - tests/test_app.py: replace quantile tests with MAE-based float reward tests - data/reward_stats.json: deleted (no longer needed) - README.md: update reward table and data generation instructions Made-with: Cursor
This config should not be part of the rdkit_chemistry MR. Keeping locally for testing only. Made-with: Cursor
Add ns_tools wrapper config for sandboxed Python/RDKit execution, rename agent to rdkit_chemistry_agent, and rewrite README to document both direct and mcp-python methods. Made-with: Cursor
Replace example rows with new ChEMBL molecules covering all five property types for both direct and mcp-python methods. Add use_box_format boolean flag to each row; rows with boxed prompts set it to true. Made-with: Cursor
Support use_box_format flag on dataset rows: when true, extract the
predicted value exclusively from \boxed{...} expressions (returning
reward 0 if absent). When false, the existing permissive cascade is
used unchanged.
Also add compute_metrics() for per-method/property-type breakdowns
and update get_key_metrics() to surface method-level stats.
Made-with: Cursor
…erver Introduce sandbox_launcher.py that starts a local nemo_skills sandbox process directly from setup_webserver(), removing the need for external sandbox orchestration in sbatch/interactive scripts. Configured via SANDBOX_VENV_PATH and SANDBOX_DISCOVERY_PATH env vars, with fallback to no-op when unset. Made-with: Cursor
Made-with: Cursor
…chemistry' Update to RDKit reward function (optional) See merge request bxyu/nemo-gym!240
jubick1337
reviewed
Mar 21, 2026
…om/bxyu/nemo-gym into dane/rdkit-chemistry
Signed-off-by: Brian Yu <bxyu@nvidia.com>
Signed-off-by: Brian Yu <bxyu@nvidia.com>
Restore the default request timeout and remove the ad hoc retry path so rollout collection matches the mainline nemo_gym behavior again. Signed-off-by: Dane Corneil <dane.corneil@gretel.ai>
Signed-off-by: Brian Yu <bxyu@nvidia.com> (cherry picked from commit 63c184e)
Record the RDKit Nano training workflow changes from the submodule so this branch points at the EOS training scripts and data-preparation flow. Signed-off-by: Dane Corneil <dane.corneil@gretel.ai>
3 tasks
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
4 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.
Implements a Nemo-Gym resources server for verifiable chemistry question answering, with and without tools (python + rdkit). The agent receives a natural-language chemistry question + SMILES molecule and must respond with a number or binary flag. The reward signal is generated from rdkit code that calculates deterministic numerical or boolean properties of SMILES: exact-match for integer/bool/fragment properties;
1/(1 + e)for continuous float properties.in the no-tools scenario, the goal is to train the model to reason about molecular properties directly and improve basic chemistry skills. In the tools condition, the goal is to train the model to use rdkit to evaluate molecular properties, when it is available.