Fine-tune MiniCPM5-1B with bare-metal TRL + PEFT, including assistant-only loss via a chat-template patch. Use when the user wants minimal Python, no YAML, full control, or asks for "TRL", "SFTTrainer", "PEFT", "LoraConfig", "assistant_only_loss".

Fine-tune MiniCPM5-1B with ms-swift (ModelScope's SFT / DPO / KTO / ORPO toolkit). Use when the user mentions "ms-swift", "swift sft", "swift rlhf", or wants ModelScope-native training. The two mandatory flags `--model_type llama --template chatml` are baked in.

Run MiniCPM5-1B with Hugging Face Transformers for one-shot Python generation on GPU (bfloat16) or CPU (float32). Use when the user wants a quick Python script, no server, no extra deps, or asks for "Transformers", "AutoModelForCausalLM", "model.generate" with MiniCPM5.

Serve MiniCPM5-1B via SGLang as an OpenAI-compatible HTTP server with RadixAttention prefix cache and built-in MiniCPM5 tool-call parsing. Use when the user asks for "SGLang", "RadixAttention", "prefix cache", batch evaluation, tool calling, or wants a high-concurrency NVIDIA-GPU server alternative to vLLM.

Run MiniCPM5-1B natively on Apple Silicon with Apple's MLX framework. Use when the user has an Apple Silicon Mac and asks for "MLX", "mlx_lm", "mlx_lm.convert", "mlx_lm.generate", or wants the fastest path on Apple Silicon.

Run MiniCPM5-1B in LM Studio (desktop GUI) using either the GGUF runtime (cross-platform) or the MLX runtime (Apple Silicon, faster). Includes OpenAI-compatible local server. Use when the user mentions "LM Studio", desktop GUI inference, "lms" CLI, or wants a no-code chat UI for MiniCPM5.

Run MiniCPM5-1B with llama.cpp using the released GGUF artifacts (F16 / Q8_0 / Q4_K_M). Use when the user wants CPU-only / consumer-GPU / cross-platform native deployment, asks for "llama.cpp", "llama-cli", "llama-server", "GGUF", or has no Python available.

Train or fine-tune sentence-transformers models across `SentenceTransformer` (bi-encoder, dense or static embedding model for retrieval, similarity, clustering, classification, paraphrase mining, dedup, multimodal), `CrossEncoder` (reranker, pair scoring for two-stage retrieval / pair classification), `SparseEncoder` (SPLADE, sparse embedding model for learned-sparse retrieval), and `MultiVectorEncoder` (ColBERT / late-interaction, per-token embeddings scored with MaxSim). Covers loss selection, hard-negative mining, evaluators, distillation, LoRA, Matryoshka, and Hugging Face Hub publishing. Use for any sentence-transformers training task.

Fine-tunes and evaluates OpenVLA-OFT and OpenVLA-OFT+ policies for robot action generation with continuous action heads, LoRA adaptation, and FiLM conditioning on LIBERO simulation and ALOHA real-world setups. Use when reproducing OpenVLA-OFT paper results, training custom VLA action heads (L1 or diffusion), deploying server-client inference for ALOHA, or debugging normalization, LoRA merge, and cross-GPU issues.

Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning neural networks, debugging loss spikes or OOM, choosing architectures, or optimizing GPU throughput.

Use this skill when the user asks to publish, upload, sync, pull, push, diff, or maintain a MoviePilot local plugin in a GitHub repository. Covers using the configured MoviePilot GitHub token, PLUGIN_LOCAL_REPO_PATHS local plugin repositories, package.json/package.v2.json metadata, plugins/plugins.v2 layouts, safe file exclusion, diff preview before publishing, incremental GitHub Contents API updates, and syncing local plugin changes back from GitHub. Includes asking whether to use an existing repository or create a new public repository when no target repository is available. Also use for Chinese requests mentioning 插件发布, 插件维护, 推送插件到 GitHub, 从 GitHub 拉取插件, 同步本地插件仓库, 增量发布插件, 插件仓库维护.

Use this skill when the user asks to operate MoviePilot through the local `moviepilot tool` MCP CLI for normal product workflows: media search, torrent search, downloads, subscriptions, downloader tasks, library checks, sites, schedulers, workflows, and messages. Prefer dedicated skills for slash command dispatch, manual file organization or failed transfer retry, direct REST API calls, direct database SQL, browser operations, and restart/upgrade.

KFL2 (Kubeshark Filter Language) reference. This skill MUST be loaded before writing, constructing, or suggesting any KFL filter expression. KFL is statically typed — incorrect field names or syntax will fail silently or error. Do not guess at KFL syntax without this skill loaded. Trigger on any mention of KFL, CEL filters, traffic filtering, display filters, query syntax, filter expressions, write a filter, construct a query, build a KFL, create a filter expression, "how do I filter", "show me only", "find traffic where", protocol-specific queries (HTTP status codes, DNS lookups, Redis commands, Kafka topics), Kubernetes-aware filtering (by namespace, pod, service, label, annotation), L4 connection/flow filters, time-based queries, or any request to slice/search/narrow network traffic in Kubeshark. Also trigger when other skills need to construct filters — KFL is the query language for all Kubeshark traffic analysis.

SERP-based semantic topic clustering for content architecture planning. Groups keywords by actual Google SERP overlap (not text similarity), designs hub-and-spoke content clusters with internal link matrices, and generates interactive visualizations. Optionally executes content creation if claude-blog is installed. Use when user says "topic cluster", "content cluster", "semantic clustering", "pillar page", "hub and spoke", "content architecture", "keyword grouping", or "cluster plan".

Use when the user wants to understand how a sales or customer-facing rep's call behavior is changing over time. Produce an evidence-backed trend readout with improvements, regressions, stable patterns, coaching actions, and calls to re-listen.

Understand and customize the local Trellis architecture inside a user project. Use when modifying .trellis plus platform hooks, settings, agents, skills, commands, prompts, workflows, the channel runtime (trellis channel), bundled runtime agents under .trellis/agents/, selectable workflow templates, registry-backed spec refresh, cross-session memory (trellis mem) generated by trellis init, or AI-facing bundled skills (trellis-channel, trellis-session-insight, trellis-spec-bootstrap) and bundled-skill auto-dispatch flow.

Generate, edit, and read PowerPoint presentations. Create from scratch with PptxGenJS (cover, TOC, content, section divider, summary slides), edit existing PPTX via XML workflows, or extract text with markitdown. Triggers: PPT, PPTX, PowerPoint, presentation, slide, deck, slides.

Open, create, read, analyze, edit, or validate Excel/spreadsheet files (.xlsx, .xlsm, .csv, .tsv). Use when the user asks to create, build, modify, analyze, read, validate, or format any Excel spreadsheet, financial model, pivot table, or tabular data file. Covers: creating new xlsx from scratch, reading and analyzing existing files, editing existing xlsx with zero format loss, formula recalculation and validation, and applying professional financial formatting standards. Triggers on 'spreadsheet', 'Excel', '.xlsx', '.csv', 'pivot table', 'financial model', 'formula', or any request to produce tabular data in Excel format.

Orchestrate the full metabolic flux analysis pipeline from model loading to phenotype prediction and publication figures. Triggers when the user provides an organism name, BIGG model ID, or custom reaction list and wants end-to-end metabolic modelling run automatically.

Run ApiChief in the EF Core repo to emit baselines, summaries, deltas, review files, or breaking-change checks. Use when refreshing `*.baseline.json`, preparing API review artifacts, or validating API changes.