Results for “pretraining”

Pretrains large language models at scale using PyTorch-native torchtitan with 4D parallelism, Float8, and distributed checkpointing.

Train, evaluate, export, and run inference for Masked Auto-Encoder (MAE) models for self-supervised pretraining and fine-tuning of visual representations.

Collects launch inputs and runs preflight checks before executing TAO workflows such as AutoML, training, evaluation, inference, export, TensorRT engine generation, or DEFT jobs on supported platforms.

Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. Use when training large MoE models with FP8/INT4, needing train-inference alignment, or requiring speculative RL for maximum throughput.

Evaluates a local pre-trade checklist before manual order entry, blocking planless, oversized, revenge-risk, market-regime-blocked, or circuit-breaker-blocked entries while journaling the decision for later review.

Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. Use when training large MoE models with FP8/INT4, needing train-inference alignment, or requiring speculative RL for maximum throughput.

Use when a commit's changes are finished and about to be presented for review. Runs the checks that catch the objections the user would otherwise raise, before they spend time reviewing.

Load this skill when the user says "prepare base", "prepare-base", or "/prepare-base". Prunes remote tracking references, checks out a base branch, pulls latest, and cleans up local-only feature branches after confirmation.

Kinetics-400: A Large Video Understanding Dataset

Internal technical readiness drill assistant for release and handoff.

在启动高风险项目或重要决策前，需要预先识别潜在失败点以增强心理韧性和方案鲁棒性时

Train large-scale MoE models with FP8/INT4 low-precision RL, speculative decoding, and train-inference alignment using the miles framework.

Run a pre-mortem — imagine the project has already failed one year from now and work backward to find the root causes before they happen. Load when the user asks for a pre-mortem, wants to imagine failure before committing, asks "what could go wrong before we start", "assume this fails — why", "risk analysis before launch", "what kills this project", "what could go wrong", or when deep-thinking diagnoses a pre-mortem frame. Based on Gary Klein's prospective hindsight method, which surfaces failure causes more effectively than forward risk analysis. Most useful right before a major commitment.

Post-training model validation workflow: gating, backtesting, walk-forward validation, deployment decisions. Trigger after GPU training completes.

Visual Prompt Tuning

Runs before every git commit on heleni-memory. Checks: no credentials in tracked files, no oversized files, valid SKILL.md frontmatter for any new/modified skill. Triggers: "commit", "before push", "validate repo", or auto-runs as a git pre-commit hook.

Pretrains large language models from scratch using PyTorch-native distributed training with 4D parallelism (FSDP2, TP, PP, CP) and Float8 support on H100 GPUs.

在启动一个重要项目或关键行动之前，用于主动识别和防范风险

Use this skill whenever a developer wants to turn an idea into a complete, production-ready, end-to-end system plan BEFORE writing any code. Trigger on 'plan this system', 'design the architecture for', 'help me blueprint', 'deep plan for X', 'break this idea into components', 'expand into action points', 'full implementation plan', or when the user pastes a project idea wanting architecture, components, pipelines, and file-level execution mapped out. Casual phrasing also triggers: 'help me think this through end-to-end', 'plan before coding'. Also covers living-workspace patterns: self-improving knowledge bases, reflection loops with auditor agents, four-agent teams, schema-as-code, wiki health scoring. **Resume triggers**: 'where did we leave off', 'continue this plan', 'resume my blueprint' - rehydrates state from disk via SESSION.md/NEXT.md/decisions/. Web research is mandatory every invocation.

Records Clean Room initialization preferences, separated artifact locations, model policy, schema profile, and clean-safe rule defaults before a clean-room run starts or resumes.

Train machine learning models end-to-end, covering data loading, preprocessing, architecture selection, training loops, validation, and checkpointing. Use when the user requests model training or provides relevant inputs for this workflow.

VILA: On Pre-training for Visual Language Models

Fine-tune a TAO model with standard supervised training, evaluation, and export, with AutoML bypass and platform-specific credential intake.

Train or fine-tune sentence-transformers models for retrieval, similarity, clustering, classification, and reranking, with support for bi-encoders, cross-encoders, and sparse encoders.

Orchestrates the full-site migration preparation cascade — inventory crawl, page-type and module catalog confirmation, archetype prototyping, and asset preparation — with confirmation gates between phases.

Set up Husky pre-commit hooks with lint-staged (Prettier), type checking, and tests in the current repo. Use when user wants to add pre-commit hooks, set up Husky, configure lint-staged, or add commit-time formatting/typechecking/testing.

Complete meeting lifecycle skill — preparation before and follow-up after any professional or personal meeting. Use this skill whenever the user mentions an upcoming meeting, wants to prepare for a meeting, needs an agenda, wants to brief themselves before talking to someone, needs to write up meeting notes, wants to send a follow-up email after a meeting, needs to capture action items, or says anything like "I have a meeting with...", "I just got out of a meeting", "help me prepare for...", "I need to follow up on...", or "what should I ask in this meeting." Also trigger for specific meeting types: job interviews, performance reviews, 1:1s, sales or client calls, and difficult conversations. Both modes are covered: PRE-meeting preparation and POST-meeting follow-up. Trigger proactively even when the user doesn't use the word "meeting" — if they describe a situation where they're about to talk to someone important, or just finished such a conversation and need to capture it, this skill applies.

Walk into the negotiation with a plan, not a hope. Sets your walk-away and your target, builds the concession ladder (what you trade and what you get for it), maps their likely asks, and drafts the responses that protect price and terms. Built for B2B sales teams, customizable to your CRM and your deal process. Trigger on "prep me for this negotiation", "the prospect wants a discount", "what should I concede", "build my concession ladder", "hold price on a new deal", or any pre-negotiation planning.

Chameleon: Mixed-Modal Early-Fusion Foundation Models

Runs a read-only pre-deploy verification gate covering git status, builds, tests, migrations, secrets, and commit conventions, then reports READY or NOT READY.

Facilitate a structured sprint retrospective to surface insights and produce actionable improvements with owners and deadlines.

Defines scope, objectives, rules of engagement, threat model selection, and operational timelines for red team engagements before any offensive testing begins.