v0.9.1 — First Continuous Training on 7B

What's New

First Successful Continuous Training Run

The continuous learning pipeline completed its first end-to-end training cycle on the full 7B model — 32 filtered samples trained, exported to GGUF, and hot-swapped into production.

OOM Batch Size Fix

The auto batch size calculator was calibrated for Qwen2.5-0.5B (40MB/sample), causing immediate OOM on the 7B model at batch_size=32. Fixed to 2GB/sample, giving batch_size ~6 which fits comfortably in 32GB VRAM.

Semantic Versioning for Model Deploys

The learning daemon now tracks model versions via MODEL_VERSION file, bumps patch version on each successful deploy, and publishes model_version in get_stats output.

Changes

• train_lora.py — Fixed batch size estimate (40MB → 2GB per sample)
• continuous_learning.py — Added _get_model_version(), _bump_model_version(), model_version in published stats
• server.py — model_version field in get_stats response
• MODEL_VERSION — New file tracking semantic version
• Training threshold lowered from 100 → 50 samples

Stats

Metric	Value
Model version	0.9.1
Training samples	32
Full cycle time	~3 min (train + export + deploy)
VRAM headroom	11 GB free (dual LLM)
RAG hit rate	82.8%
Learning queue	51 raw / 32 filtered

Full Changelog: v0.9.0...v0.9.1