014ec8bd5c
- Training data pipeline: convert, export, extract, load-to-db - Infra tooling: infra-audit, infra-gitea-link - RAG pipeline: rag-ingest, rag-query - Fine-tuning: finetune-lora, overnight-qwen3, install-unsloth - Transcripts: export-transcripts - Updated README with script index and token reduction strategy
grace-scripts
Grace AI helper scripts — token-efficient homelab automation.
These scripts replace inline shell construction in AI sessions, reducing token usage and making repeated tasks reproducible and auditable.
Scripts
| Script | Purpose |
|---|---|
convert-training-data.py |
Convert raw session exports to unsloth/axolotl-ready JSONL |
export-training-data.py |
Export training data from OpenClaw sessions |
export-transcripts.py |
Export session transcripts incrementally |
extract-sessions.py |
Extract and parse session data |
finetune-lora.py |
Run LoRA fine-tuning on local Qwen3-8B |
infra-audit.py |
Audit homelab infra state and sync to MongoDB |
infra-gitea-link.py |
Link infra components to Gitea issues/repos |
load-transcripts-to-db.py |
Sync transcripts to PostgreSQL on DB VM |
overnight-qwen3.py |
Overnight batch inference job |
overnight-research.sh |
Overnight research and summarization |
rag-ingest.py |
Ingest documents into Qdrant for RAG |
rag-query.py |
Query Qdrant RAG index |
install-unsloth.sh |
Install unsloth fine-tuning framework |
Credential Protocol
All scripts pull credentials from Infisical. Never hardcode secrets. IPs and service URLs reference TOOLS.md / BOOTSTRAP.md in the Grace workspace.
Token Reduction Strategy
Each script replaces multi-step inline shell that Grace would otherwise construct token-by-token in every session. One exec call to a script = ~10-50x fewer tokens than building the equivalent inline.
Issues & Improvements
Use Gitea issues on this repo to track bugs, improvements, and new script requests.