Merge pull request #9 from evasnow1992/contrastive_kermt

KERMT v2.0.0 release: Contrastive KERMT + agent skills

Author: sveccham

.dockerignore

@@ -0,0 +1,31 @@
+# Large model checkpoints — never needed in the image; mount at runtime via
+# kermt_container.sh --ckpt instead.
+model/
+
+# Git metadata.
+.git/
+.gitignore
+.gitmodules
+
+# Python caches.
+**/__pycache__/
+**/*.pyc
+**/*.pyo
+
+# Runtime artifacts created by agent workflows.
+runs/
+
+# Agent/editor metadata that should not ride along inside the container.
+.claude/
+CLAUDE.md
+CODEX_MEMORY.md
+.vscode/
+.idea/
+
+# OS-level noise.
+.DS_Store
+Thumbs.db
+
+# Dockerfile is read by `docker build` directly; no need to re-COPY it into /code.
+Dockerfile
+.dockerignore

.gitignore

@@ -137,3 +137,10 @@ dmypy.json
 
 # created pickles
 exampledata/finetune/*.npz
+
+# Local-only datasets and trained models (not intended for upstream)
+model/
+tests/data/Biogen_for_grover/
+tests/data/ChEMBL_MT/
+tests/data/ZINC15_ChEMBL/
+tests/data/openadmet/

AGENTS.md

@@ -0,0 +1,20 @@
+# Agent integrations
+
+Agent-driven workflows for KERMT live under [`agent/`](agent/README.md). The
+skills follow the [agentskills.io](https://agentskills.io) spec — one
+directory per skill at `agent/skills/<skill-name>/SKILL.md`.
+
+Available workflows:
+- `kermt-setup` — bootstrap the kermt container (run first)
+- `kermt-continue-pretrain` — continue pretraining from an existing KERMT ckpt
+- `kermt-pretrain-scratch` — pretrain from scratch on a custom corpus
+- `kermt-add-cmim-pretrain` — add a cMIM decoder to a grover_base ckpt
+- `kermt-finetune` — finetune a pretrained ckpt on a labeled CSV
+- `kermt-infer` — predictions with a finetuned ckpt
+- `kermt-embed` — extract molecular embeddings from any encoder-bearing ckpt
+- `kermt-monitor` — tail logs / report progress for a detached run
+
+See [`agent/README.md`](agent/README.md) for the full guide: hardware
+requirements per workflow, how to install the skills with Claude Code /
+Codex / Nemotron, and how to invoke the underlying scripts directly without
+an agent.

Dockerfile

@@ -42,9 +42,12 @@ RUN rm -rf /var/lib/apt/lists/*
 RUN rm -rf /softwares/miniconda3/pkgs/xz-5.6.4-h5eee18b_1/
 # RUN apt clean && apt autoremove -y
 
-# COPY code and cleanup
-COPY . /code
-RUN rm -rf /code/.git /code/.code-workspace
+# Note: the repo is NOT copied into the image. Agent skills and any equivalent
+# host workflow bind-mount the live kermt repo checkout at /workspace (see
+# agent/scripts/kermt_container.sh). Keeping the image as a pure environment
+# makes rebuilds cache-friendly (only invalidates when environment.yml changes,
+# not on every code edit) and avoids stale-code footguns where a baked-in /code
+# and the runtime bind-mount disagree.
 
 # Equivalent to `conda activate kermt`
 SHELL ["conda", "run", "--no-capture-output", "-n", "kermt", "/bin/bash", "-c"]

Pipfile

@@ -0,0 +1,22 @@
+[[source]]
+url = "https://pypi.org/simple"
+verify_ssl = true
+name = "pypi"
+
+[packages]
+descriptastorus = "==2.7.0"
+optuna = "==4.9.0"
+pip = "==26.1.2"
+protobuf = "==6.33.5"
+pytest = "==9.0.3"
+rdkit = "==2025.9.1"
+scikit-learn = "==1.9.0"
+setuptools = "==81.0.0"
+tensorboard = "==2.20.0"
+torch = "==2.9.1"
+wandb = "==0.26.1"
+
+[dev-packages]
+
+[requires]
+python_version = "3.11"