BibleLM v1.1.0 — GPU Memory Guard + Ask Tab

What's new in v1.1.0

GPU memory guard

The Training tab now shows a live Est. VRAM bar that updates as you adjust any config slider. It estimates memory needed for model weights, gradients, Adam optimizer states, and attention activation maps — and blocks Start Training if the estimate exceeds 90% of your GPU's capacity. No more out-of-memory crashes from misconfigured runs.

The biggest VRAM consumers: batch_size × context_len² (attention maps) and d_model (parameter count). If you're near the limit, reduce batch_size first.

Ask tab — exact KJV search engine

Deterministic search over all 31,102 KJV verses. Understands natural-language queries:

• John 3:16 — look up any verse reference
• how many verses contain "love" — word occurrence count
• most common word in Psalms — frequency stats
• longest verse in the Bible — length queries
• how many chapters in Revelation — book stats
• Free-text search across all verses

Training UX

• Defaults button — reset config sliders to the recommended starting point
• Device card — shows CUDA (GPU) or CPU once training starts
• Waiting overlay — spinner while the Python process spins up before the first metric arrives

Chart fix

Y-axis labels now adapt to available chart height — no more overlapping labels when the chart is tall or narrow.

---

Installation

Python and PyTorch are not included in the installer. You must install them yourself before the app can train or generate. See the README for setup instructions.

1. Run BibleLM_1.1.0_x64-setup.exe
2. Install Python 3.13 and create a .venv at the repo root (see README)
3. Install PyTorch (CUDA build recommended for speed)
4. Launch BibleLM

Building from source is recommended if you haven't used v1.0.0 — the installer bakes in the build-time source path, so it works best on the machine it was compiled on.

---

Pre-trained model artifacts

The ckpt.pt / model.onnx / tokenizer.json files below are the same model from v1.0.0:

• Architecture: 256 d_model · 4 layers · 4 heads · 256 context → 3.2M parameters
• Training: 15 epochs on KJV, val loss 1.19
• Hardware: RTX 4070 SUPER — ~290K tok/s

Drop ckpt.pt + tokenizer.json into runs/current/ to use Inference without training.