training-data-bridge

name	training-data-bridge
description	Extract training signals from Fae's memory — SFT examples, DPO correction pairs, engagement scores, and interest-weighted sampling.
tags	["training","self-improvement","memory"]
metadata	{"author":"fae","version":"1.0"}

Training Data Bridge

Converts Fae's memory into weighted training data for personal model improvement.

Available Scripts

• build_dataset: Complete pipeline — exports episodes, extracts corrections, computes weights, validates
• extract_corrections: Detect implicit and explicit user corrections → DPO training pairs

What It Extracts

Memory Kind	Training Signal
.episode	Raw SFT examples (conversation turns)
.profile	Style preferences (response format shaping)
.interest	Topic weighting (upweight relevant domains)

Correction Detection

Patterns detected for DPO pair generation:

• Explicit rephrase: "Too long", "Shorter please"
• "No, I meant..." corrections
• Silent abandonment → retry (5+ min gap then rephrase)

Output

Generates files in ~/Library/Application Support/fae/training/data/:

• train.jsonl — Weighted SFT training examples (80% split)
• valid.jsonl — Validation examples (20% split)
• dpo_pairs.jsonl — Correction-based preference pairs
• meta.json — Dataset statistics and quality report