ADR-129: RuvLTRA GCloud Training Pipeline with TurboQuant Optimization

Open

#310 opened on Mar 28, 2026 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

enhancement

Dominant language: Rust
Stars: 4.5k
Forks: 603
Avg merge: 23h 32m
Merged PRs (30d): 59

Description

Overview

Implement the full training pipeline from ADR-129 to retrain RuvLTRA models with TurboQuant KV-cache profiling on Google Cloud.

Phases

Phase 1: imatrix Recalibration + TurboQuant KV Profiling (Week 1)

Build gcr.io/ruv-dev/ruvltra-training:latest Docker image
Run imatrix recalibration with code-focused calibration data
Generate .turboquant.json sidecar profiles per model
Benchmark recalibrated GGUFs vs baseline (ablation run B)

Phase 2: WET-Augmented LoRA Fine-Tuning (Week 2-3)

Export brain memories + WET data as training corpus
Run eval contamination check (13-gram overlap)
Validate dataset governance (schema, dedup, quality scores)
Run LoRA SFT on Vertex AI A100 (ablation run C)
Run DPO training (ablation run D)

Phase 3: Benchmarking & Validation (Week 3-4)

Run full ablation matrix (runs A-E)
Evaluate all 7 release gates (G1-G7)
Produce contamination report + ablation report
Automate via scripts/training/release_gate.py

Phase 4: Publishing (Week 4)

Produce GGUF variants + .turboquant.json sidecars
Publish to HuggingFace (all 4 models)
Update model cards with benchmark results
Update ruvllm registry with checksums
Publish ruvllm and @ruvector/ruvllm with sidecar loading
Set up weekly benchmark scheduler job

Release Gates (G1-G7)

Gate	Criterion
G1	HumanEval pass@1 ≥ 45% (0.5B) / ≥ 55% (3B)
G2	Routing accuracy ≥ 80% (no regression)
G3	Wikitext-2 PPL increase < 5%
G4	TurboQuant ≥ 8x compression, PPL delta < 1%
G5	Long context PPL < 20 at 16K tokens
G6	Zero eval contamination
G7	Inference ≥ 80 tok/s (0.5B) / ≥ 40 (3B)

Infrastructure

Compute: L4 GPU (Cloud Run Jobs) + A100-80GB (Vertex AI)
Data: Brain memories (3,870+), WET corpus, Claude Flow routing (2,700+), ADR corpus (129 docs)
Estimated cost: ~$70-210 (experimental compute)

Files Created

scripts/training/release_gate.py — Automated ship/no-ship checker
scripts/training/export_training_data.py — Dataset export with governance
scripts/training/contamination_check.py — Eval contamination detection
scripts/training/Dockerfile — Training image
scripts/training/deploy_training.sh — Cloud Run job creation
scripts/training/run_calibration.py — Phase 1 entry point
scripts/training/run_sft.py — Phase 2 entry point
crates/ruvllm/src/quantize/turboquant_profile.rs — Sidecar config loading

ADR-129: docs/adr/ADR-129-ruvltra-gcloud-training-turboquant.md
TurboQuant: crates/ruvllm/src/quantize/turbo_quant.rs
Models: ruv/ruvltra-claude-code (7,615 downloads)

🤖 Generated with claude-flow

Contributor guide

No contributing guide indexed for this repository

First steps

Read the whole issue, then the project's contributing guide.
Comment on the issue to say you are picking it up — it saves two people doing the same work.
Fork the repository and make your change on a branch.
Open a pull request that references the issue number.

Research direction

Start with docs/adr/ADR-129-ruvltra-gcloud-training-turboquant.md, then inspect the listed entry points under scripts/training/ and crates/ruvllm/src/quantize/turboquant_profile.rs. The work spans four phases, seven release gates, cloud training, benchmarking, publishing, and scheduling; it is done when those phases and gates are completed and the model, registry, and package updates are published.

Written by the indexing model from the issue text.

Assessment

Tech stack: docker, gcp, huggingface, python, rust
Domain: ai, cloud, devops, machine-learning
Issue type: Feature
Difficulty: 5/5
Estimated time: Over a week
Activity status: Quiet
Clarity: Needs clarification
Newbie friendliness: 25/100

A short digest of beginner-friendly GitHub issues.