Pipelines
This project now supports reusable pipelines for: - indexing all docs into JSON - building per-training corpora from those docs - creating separate model bundles - training new bundles from scratch - fine-tuning existing bundles with new/updated docs - asking questions against a specific bundle with
SmallLanguageModel Pipelines
This project now supports reusable pipelines for:
0) Environment
Use .venv/bin/python in the commands below if your system Python does not have the required packages.
2) Build JSON Index + JSONL Content From Scan
This index now includes rich metadata per file: path, hashes, bytes, chars, lines, words, timestamps, and token counts (tokenswhitespace, tokensestimate4chars). If you want tokenizer-accurate token counts, add --tokenizerfile artifacts/models/<model/tokenizer/tokenizer.json.
3.5) Build Retrieval Chunks (Recommended)
buildchunkindex.py does not train the model; it prepares retrieval chunks used by RAG.
5) Initialize a New Model Bundle
Creates an isolated bundle under artifacts/models/<modelname.
8) Clone Existing Model Into New Version (Recommended)
This keeps the old model frozen and trains a new child version.
9) Fine-Tune Existing Bundle In Place (Optional)
Rebuild index first, then changed-only corpus: