Skip to content

fix(neural): enhance multivariable partial/gradient training, constrained decoding, and update checkpoint - #66

Merged
chaudhryumer merged 6 commits into
QuantumLogicsLabs:mainfrom
umer12x:DEV2_task
Oct 18, 2026
Merged

chaudhryumer merged 6 commits into
QuantumLogicsLabs:mainfrom
umer12x:DEV2_task

Conversation

@umer12x

@umer12x umer12x commented Oct 17, 2026 •

Copy link
Copy Markdown
Contributor

Overview

This pull request introduces critical improvements to neural calculus solving, specifically targeting accuracy and structural validity on multivariable partial derivatives and gradient vector outputs. It includes data generation upweighting, decoder alignment fixes, constrained beam search decoding, and updated model checkpoints with evaluation documentation.


Key Changes

  1. Scheduled Sampling Alignment Fix (train.py)

    • Corrected decoder input alignment during scheduled sampling where sampled predictions replace teacher-forcing tokens, shifting predictions by one position to match sequential autoregressive targets.
  2. Multivariable & Gradient Dataset Rebalancing (problem_generator.py)

    • Upweighted the generation ratios of partial derivative and multivariable gradient expressions across all difficulty tiers.
    • Re-generated the 75k sample dataset splits (data/splits/train.jsonl, val.jsonl, test.jsonl, slang_dataset.jsonl) with balanced variable distributions across single-variable and multi-variable domains.
  3. Constrained Beam Search Decoding (inference/, deployment/)

    • Added variable-consistency constraints during beam search decoding for multivariable expressions.
    • Enforced syntax validity and gradient vector dimension constraints in:
      • inference/beam_search.py
      • inference/solve.py
      • deployment/onnx_beam_search.py
      • deployment/onnx_solve.py
  4. Inference & Evaluation Hyperparameters (eval/)

    • Increased default beam size to 5 and maximum sequence length to 80 in evaluation scripts for complete prefix-expression generation.
    • Added scheduled sampling ablation runner script (scripts/ablation_scheduled_sampling.py).
  5. Updated Checkpoint & Benchmarks (checkpoints/, docs/)

    • Updated trained weights in checkpoints/final/best.pt.
    • Added updated evaluation logs and benchmark metrics in docs/EVAL_AFTER_FIX.md and docs/eval_after_fix.json.

Verification

  • Tested beam search decoding across single and multivariable problem sets.
  • Verified syntax validity and mathematical equivalence checks using SymPy verification harness.

@vercel

vercel Bot commented Oct 17, 2026

Copy link
Copy Markdown

@umer12x is attempting to deploy a commit to the seno-quantum-coder's projects Team on Vercel.

A member of the Team first needs to authorize it.

@umer12x umer12x changed the title Dev2 task fix(neural): enhance multivariable partial/gradient training, constrained decoding, and update checkpoint Oct 18, 2026
@chaudhryumer
chaudhryumer merged commit 85725da into QuantumLogicsLabs:main Oct 18, 2026
1 of 5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants