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 skills add BillusA1111/billus-model-skill-library --skill billus-model-opsgit clone --depth 1 https://github.com/BillusA1111/billus-model-skill-libraryWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/billusa1111/billus-model-skill-library/billus-model-ops)<a href="https://agentmods.dev/skills/billusa1111/billus-model-skill-library/billus-model-ops"><img src="https://agentmods.dev/badge/skills/billusa1111/billus-model-skill-library/billus-model-ops/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/billusa1111/billus-model-skill-library/billus-model-ops"><img src="https://agentmods.dev/badge/skills/billusa1111/billus-model-skill-library/billus-model-ops.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What 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 5.1 | $0.00113 | $0.01254 |
| Opus 5 | $0.00056 | $0.00627 |
| Sonnet 5 | $0.00023 | $0.00251 |
| Haiku 4.5 | $0.00011 | $0.00125 |
Grade A, and why
billus-model-ops 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 10d 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 — 96 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Billus Model Ops
Overview
Use this skill for the model repos you touch every day: Hugging Face training stacks, DeepSpeed or FSDP scale-up flows, PEFT or QLoRA adaptation, VL wiring, and diffusion-style image generation or editing systems. Start from repository evidence, identify the active framework path, make the smallest viable change, and leave behind validation notes another engineer can trust.
Quick Start
- Identify the framework stack.
- Run
python scripts/detect_training_stack.py <repo-root>when the repo is unfamiliar. - Confirm the actual launch path with
rg -n "accelerate launch|deepspeed|torchrun|Trainer\\(|DiffusionPipeline|LoraConfig|vision_tower|projector|ControlNet" <repo-root>.
- Run
- Classify the task.
- Trainer, scheduler, batch size, optimizer, data path: read
references/frameworks-hf-accelerate-deepspeed.md - LoRA, QLoRA, adapter merge, low-bit finetune: read
references/frameworks-peft-and-lowbit.md - LLaVA, Qwen-VL, InternVL, projector, image tokens, visual encoder: read
references/frameworks-vl-stacks.md - diffusers, image edit, inpainting, ControlNet, IP-Adapter, scheduler, VAE: read
references/frameworks-diffusers-image.md - Before final delivery, always read
references/validation-and-release.md
- Trainer, scheduler, batch size, optimizer, data path: read
- Keep the change bounded.
- Prefer modifying the existing config or launch surface instead of introducing duplicate knobs.
- Keep baseline seeds, datasets, eval prompts, and sample grids stable unless the task explicitly changes them.
- Validate before handoff.
- Use the lightest repo-native smoke test that still exercises the changed path.
- Summarize what was validated, what still needs a full run, and the main rollback lever.
Workflow
1. Map the active stack
Identify:
- Trainer style: Hugging Face Trainer, raw PyTorch loop, Accelerate custom loop, DeepSpeed launcher, FSDP wrapper, diffusers trainer, or mixed custom code
- Config surface: argparse, dataclass, YAML, JSON, Hydra, shell scripts, environment variables
- Model family: decoder-only LLM, encoder-decoder, VL with vision tower plus projector, diffusion or flow image model
- Runtime topology: single GPU, DDP, ZeRO, FSDP, CPU offload, mixed precision, low-bit adapters
What ships with it
9 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- agents/openai.yaml 297 B
- references/frameworks-diffusers-image.md 1.5 KB
- references/frameworks-hf-accelerate-deepspeed.md 2.1 KB
- references/frameworks-peft-and-lowbit.md 1.5 KB
- references/frameworks-vl-stacks.md 1.5 KB
- references/validation-and-release.md 970 B
- scripts/detect_training_stack.py 3.5 KB runs code
- scripts/new_experiment_note.py 2.3 KB runs code
- scripts/summarize_training_log.py 6.0 KB runs code
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.
- 10d ago First seen · 96 lines · 113 tokens per session scan A 194b6a58b303
billus-model-ops is a skill published in the GitHub repository BillusA1111/billus-model-skill-library (11 stars, last pushed 6mo ago), licensed Apache-2.0. It adds 113 tokens to every session and 1,254 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 2026-08-30.
Other skills, from other repositories
huggingface-accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
axolotl
Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support.
huggingface-accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
axolotl
Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support.
huggingface-accelerate
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
axolotl
Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support.