dotfiles/claude/memory/training-pipeline.md at 3dff1a93d9aa00d26fa4692a25436ef821d50115

Files

T

David F GliddenandClaude Opus 4.8 3f9a89b00c chore(memory): Basic Memory trial begins — sync normalization baseline (283 files)

Basic Memory v0.21.6 first sync over the live memory dir (steward-authorized
live-dir trial, Option A 2026-06-06): adds permalink: to frontmatter, refolds
long YAML description lines, strips final newlines. Bodies untouched —
verified via full diff classification. From this commit forward, any diff in
claude/memory shows only what Basic Memory or the session writes.

Trial design: MemPalace untouched as incumbent; git status check on this dir
at every wrap; end-of-day evaluation (recall quality, sync robustness,
rebuild-from-files, malformed-file behavior).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-06 09:52:17 +02:00

3.4 KiB

Raw Blame History

name, description, type, permalink

name	description	type	permalink
training-pipeline-state	LoRA training pipeline — model trained, GGUF transfer failed over VPN. Must download on home LAN. Then rebuild BMF + reconnect MCP.	project	claude-memory/training-pipeline

Training Pipeline — State as of 2026-03-16 (evening)

MODEL TRANSFER — BLOCKED ON LAN ACCESS

6.2 GB GGUF on clasp at /tmp/lora-david/model-f16.gguf
~30 rsync attempts over VPN from Berlin failed: hash mismatches from partial corruption + 120s timeout too short for 1.4 MB/s VPN throughput
Partial file at /tmp/cm-david-model.gguf is corrupt — delete before next attempt
Script ready: /tmp/pull-model-vpn.sh (600s timeout, SSH keepalive, auto-registers in Ollama)
Action when home (2026-03-17): rm /tmp/cm-david-model.gguf && nohup /tmp/pull-model-vpn.sh > /dev/null 2>&1 &
On LAN (~200 MB/s) a clean transfer takes ~30 seconds
Log: tail -f /tmp/pull-model.log

WHEN MODEL LANDS

Verify: ollama list | grep cm-david
Rebuild BMF: cd ~/_ Dev/BetterMemories.io && npm run build
Reconnect MCP in Claude Code: /mcp
This activates Seb's fixes: teacher during ingestion (#49), chat export (#47), pairing fix (#43), graduation fallback (#45)

Seb's Fixes (pulled to main, 2026-03-16)

8dd17a0 #39 — 'auto' backend default
eca8da7 #40 — llama.cpp GGUF fallback (auto-clones)
7fec033 #43 — Pairing ID mismatch detection
422212f #45 — Graduation fallback when teacher unavailable
6c332b7 #47 — Chat format in export_training_pairs
ce03770 #49 — Teacher enrichment during connector ingestion (rate-limited)
39915cc — Flush timer fix for circle pairings

Second Training Run (COMPLETE)

14,199 deduplicated pairs from David's instance (vault + git + BMF sessions)
Qwen2.5-3B-Instruct, rank-16, 2,000 iters, 21 min on M4 Pro
Output: /tmp/lora-david/ on clasp, registered as cm-david-all-slots:lora-v1

Clasp State (cleaned up this session)

122 stale Modelfile models removed
PATH fixed (#42): node/npm available in SSH sessions via ~/.zshenv
TypeScript rebuilt with #39 fix
mlx-fine-tune.py updated with llama.cpp fallback
BMF v0.48.0 running, healthy

Ingestion

Claude transcripts: idle / needs restart — was at 7,965/72,938 (11%) but BMF process restart lost the job
Restart after model lands + BMF rebuild: connect_service with claude_transcripts
With Seb's #49 fix + MCP reconnect, remaining transcripts WILL generate training pairs via teacher

Remaining Issues (not yet resolved by Seb)

#41 — Qwen3.5 GGUF broken (upstream, informational)
#42 — Clasp SSH PATH (we fixed manually via .zshenv)
#44 — Clasp teacher not configured (decided to skip — unnecessary token burn)
#46 — Stale models (we cleaned manually)
#48 — GGUF distribution pipeline (future feature)

Architecture Finding

Training pairs only generated via sampling/teacher provider in slot chains
All slots start SIGNALING with rules-only → no sampling → no pairs from ingestion
Seb's #49 fix adds optional teacher enrichment during ingestion (rate-limited)
#45 fix helps graduation progress when teacher unavailable

Why: Training flywheel = sovereignty. Local models that learn from your data. How to apply: Model landing → reconnect MCP → Chamber testable. Next training round after ingestion completes (more data = better model).

3.4 KiB Raw Blame History