Repository navigation
RSI: validate live LoRA or QLoRA training and paired evaluation #9
Description
Implemented
PR #14 added a fixed OpenShell QLoRA fixture-smoke command, training scripts, model and dataset provenance, and paired base/adapter evaluation wiring. Mock-queue tests cover accepted, partial, failed, and incomplete pairs. The training image was built; the available GGUF is rejected as a training checkpoint.
Remaining work
Run the documented bounded command with a pinned supported Hugging Face checkpoint and verified dataset. Record the produced adapter hash, seed, configuration, model and dataset versions, and paired base/adapter cell outcomes. Confirm that failed or partial training remains ineligible for selection. No live adapter training and paired evaluation has yet been documented.
Acceptance criteria
- A documented command can train one bounded adapter from a verified dataset.
- The produced artifact is evaluated against the base model on the same harness cells.
- Failed or partial training is recorded without making the candidate eligible for selection.
Activity
[-]RSI: connect model candidates to a real LoRA or QLoRA backend[/-][+]RSI: validate live LoRA or QLoRA training and paired evaluation[/+]w4ffl35 commentedon Oct 13, 2026
Blocker status (2026-10-13): code side is done (PR #14: QLoRA fixture-smoke, scripts, provenance, paired-eval wiring, mock-queue tests). What remains is a live GPU run: pinned supported Hugging Face checkpoint + verified dataset, bounded training command, recorded adapter hash/seed/config and paired cell outcomes. Needs GPU time and operator scheduling. Not actionable in code until the run is scheduled.
w4ffl35 commentedon Oct 14, 2026
Live validation complete (2026-10-14). Ran the documented bounded command end to end on queue
rsi-train-20261013(dedicated; the shared queue's live workers were left untouched):rsi-train-model --resume rsi-20261014015103-ca5c1dc8 --model-candidate-id qlora-qwen3-0.6b-live2 --model meta/muse-spark-1.3-contributor --base-ref HEAD --max-concurrent 2Qwen/Qwen3-0.6B-Base@da87bfb608c14b7cf20ba1ce41287e8de496c0cd(1.2GB, shad5a97e42…), verified datasetb321139e…, manifest450cf758…, cellseda08fd3…, seed 1337, 8/8 steps, adapter 9.2MB (fe13e87f…).c417ed74completed on TRAINING_GPU (RTX 5080); base eval0c41b413+ adapter eval2817c3aecompleted on CPU. Leases released.rejected, eligible=false. No improvement claimed (expected: 8 smoke steps on 3 toy rows teach nothing; the path, not the metric, is the evidence).qlora-qwen3-0.6b-smoke4and-live1(bothpartial, eligible=false) and smoke1-3 (failed, eligible=false), all recorded in the archive with failure provenance.Blockers fixed along the way (PRs #163, #164): training image needed a C toolchain for triton; guest command pointed at unreachable /opt scripts (now run from the frozen snapshot); PEFT README broke the fixed adapter set; extractor rejected git-archive directory entries. Full record in run
rsi-20261014015103-ca5c1dc8model candidateqlora-qwen3-0.6b-live2.