Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/mxslr/mlcraft/accuracy-improvement-loopnpx skills add mxslr/mlcraft --skill accuracy-improvement-loopgit clone --depth 1 https://github.com/mxslr/mlcraftWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00112 | $0.00867 |
| Opus 5 | $0.00056 | $0.00434 |
| Sonnet 5 | $0.00022 | $0.00173 |
| Haiku 4.5 | $0.00011 | $0.00087 |
Grade A, and why
accuracy-improvement-loop scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 37 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Accuracy-Improvement Loop
Do NOT throw random tricks at the model. Diagnose, research the gap, combine principled fixes, verify. This is the loop a professional runs to push toward (and recognize) the ceiling.
Step 1 - Diagnose the bottleneck (evidence, not guesses)
- Re-audit leakage FIRST (
data-rigor-and-leakage). A suspicious jump/drop is often a split problem. Fixing leakage may lower the number - that is success, because it was fake. - Learning curves: train loss down but val loss up means overfitting; both flat or low means underfitting or a data problem; unstable means an LR, batch, or BatchNorm issue.
- Error analysis: inspect the confusion matrix and the actual misclassified samples. What do the errors have in common?
- Saliency check (Grad-CAM): is the model attending to the signal or to an artifact (background, labels, muscle/border, watermark, spurious correlation)? Fix the data/preprocessing if so.
- Per-branch/per-model check: in fusion/ensembles, is one component dragging the rest down?
Step 2 - Research the specific gap
Invoke literature-review / paper-researcher targeted at the diagnosed problem (e.g., "small-lesion whole-image classification", "fusion gradient dilution", "class imbalance long-tail"), not the generic task. Extract techniques that address this failure mode.
Step 3 - Synthesize a principled combination
Combine the best-evidenced techniques that target the bottleneck - this is how you "beat existing work": a grounded combination, not one silver bullet. Common high-impact levers (pick by diagnosis):
- Data/signal: higher input resolution, better ROI/region focus, artifact removal, cleaner labels, stronger domain-appropriate augmentation, more/balanced data (incl. synthesis only if justified).
- Architecture: stronger/more-suitable backbone for this dataset (not just the trendiest); attention (CBAM/self-attention); multi-scale features; multi-view/multi-input; feature-fusion done right - gated fusion + deep supervision / auxiliary per-branch losses to beat gradient dilution when a weak branch drags a fusion.
- Training: discriminative LR, better schedule, regularization retuned to the diagnosis, longer training, EMA/SWA, self-distillation, patch-pretraining then whole-image.
- Inference/aggregation: TTA, calibrated thresholds, ensembles (soft-voting / weighted-by-val / stacking) - output-level fusion is often the cheapest real gain.
- Weakly-supervised localization (e.g., GMIC-style) when lesions/objects are tiny in a large input - often both more accurate and gives free saliency.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 37 lines · 112 tokens per session scan A 667a37b086fe
accuracy-improvement-loop is a skill published in the GitHub repository mxslr/mlcraft (8 stars, last pushed 1mo ago), licensed MIT. It adds 112 tokens to every session and 867 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
torch-performance-optimization
Optimize or review eager CPU-only Albucore PyTorch runtime paths with benchmark-backed decisions. Use when adding or changing Torch CPU kernels, Tensor/NumPy bridges, Torch backend routing, tensor layouts, allocations, threading, profiling, memory-format candidates, or Torch performance benchmarks.
ml-for-aec
Computer vision for buildings, image-to-floorplan, generative ML models, performance prediction, structural analysis ML, energy prediction, natural language to design, and point cloud ML for AEC computational design.
yolo-export
Use when exporting or deploying Ultralytics YOLO models in Platform or code — the Platform Export tab and yolo export/model.export() for ONNX, TensorRT, CoreML, OpenVINO, LiteRT, NCNN, ExecuTorch, and NPUs (RKNN, QNN, Hailo, Ascend, IMX, Axelera, DeepX), FP16/INT8 quantization, benchmarking, and non-Python runtimes.…
yolo-training
Use when training, fine-tuning, or validating Ultralytics YOLO models in Platform, cloud GPUs, or local code — model.train(), yolo train/val, remote metric streaming, epochs, batch, imgsz, devices, augmentation, multi-GPU, resumes, results, and fixing OOM, NaN loss, low mAP, or overfitting. For hyperparameter search…
yolo-datasets
Use when uploading, annotating, building, converting, analyzing, or debugging datasets in Ultralytics Platform or local YOLO — Platform dataset management and Smart Annotation, data.yaml, YOLO label .txt formats, COCO/DOTA/mask conversion, auto-labeling, splits, validation, and errors like "no labels found" or mAP…
yolo-models
Use when choosing or comparing Ultralytics models in Platform or code — picking a model family (YOLO26/YOLO11/YOLOv8, YOLO-World, YOLOE, SAM/SAM2/FastSAM, RT-DETR, YOLO-NAS), size (n/s/m/l/x), task variant (-seg, -sem, -cls, -pose, -obb, -depth), pretrained checkpoint, open-vocabulary or promptable…