Reproducibility Kit

Toward Independently Reproducible Numbers.

A planned reproducibility kit: model weights, training code, evaluation scripts, and a 10K sample compound set. Built on the principle that scientific claims should be independently reproducible. Status: most components are still in preparation (see schedule below) — the scientific pipeline container is not yet built, so the /methodology benchmark numbers are not yet reproducible by an outside auditor.

Why Open This

Trust requires verification

In drug discovery, an unverifiable benchmark number is worth less than nothing — it costs partners time and erodes trust. We expose our model so anyone can run it.

What Stays Proprietary

Target-fine-tuned weights + ROBOGATE

Pre-train weights are open. Target-specific fine-tuned weights (TYK2, CDK4, CDK6, TNIK, EGFR) stay private — this is the result of compute investment + curated data. ROBOGATE adaptive sampling is patent-protected.

License

Apache 2.0 + CC-BY-NC

Code: Apache 2.0 (commercial use OK). Sample compound dataset: CC-BY-NC 4.0 (academic use; commercial use requires partner agreement).

Kit Components

5 Deliverables

MolForge-Gen v1 (small)

Model weights

Reduced 25.4M-parameter MolForge-Gen weights (Pre-train only, no target fine-tune). Sufficient for general molecular generation reproduction. Fine-tuned target-specific weights remain proprietary.

→ PyTorch state_dict (.pt) + config.json + tokenizer vocab

PREPARING

Apache 2.0

~95 MB

Training & Evaluation Code

Source code

Full training pipeline (tokenizer, model, training loop, evaluation scripts). Excludes ROBOGATE adaptive sampling (KIPO 10-2026-0057732 protected method).

→ GitHub repo: liveplex-cpu/molforge-gen-public

PREPARING

Apache 2.0

~2 MB

Sample Compound Set (10K anonymized)

Dataset

10,000 anonymized compound IDs with full property predictions (pIC50, hERG, AMES, QED, novelty). SMILES omitted — properties only. Sufficient for benchmarking against your own pipeline.

→ CSV + JSON + property statistics report

PREPARING

CC-BY-NC 4.0

~3 MB CSV

Benchmark Suite Runner

Source code

Scripts to reproduce our TDC ADMET, GuacaMol, and TYK2 prospective screening benchmark numbers using our model. Verifies the External Benchmarks table on our /methodology page.

→ Python scripts + Docker compose for environment

PREPARING

Apache 2.0

~500 KB

Conformal Prediction Module

Source code

Standalone CP calibration module — drop-in for any sklearn / PyTorch regressor. The exact module that produces our 86.5% empirical coverage on TYK2.

→ pip-installable: pip install molforge-conformal

PREPARING

Apache 2.0

~200 KB

Release Schedule

Phased Rollout

2026 Q2

GitHub repo + tokenizer + sample compound dataset

In Progress

2026 Q3

Pre-train weights (small) + training code + benchmark runner

Planned

2026 Q4

Conformal Prediction pip package + Docker image

Planned

2027 Q1

Public TDC leaderboard submission with reproducible scripts

Planned

Get Notified on Release

Be First When the Repo Goes Public

Email contact@molforgeai.com with subject "repro-kit" to get an early-access notification when each component is released.

Request Access →