skills.rest
chengyansen-ai
chengyansen-aiCOMMUNITY
Explore all skills by this author →
GitHub

h3-multimodal-lora-training

Plan, audit, train, and validate MiniMax H3 multimodal LoRA adapters through a ten-step evidence chain.

25|1|Updated Aug 24, 2026
Search Tags:#lora#diffusion#checkpoint validation#dataset audit#minimax h3#ref2va#video training
One-click install
npx skills add https://github.com/chengyansen-ai/h3-multimodal-lora-training --skill h3-multimodal-lora-training-chengyansen-ai
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: h3-multimodal-lora-training
Source: https://github.com/chengyansen-ai/h3-multimodal-lora-training
Command: npx skills add https://github.com/chengyansen-ai/h3-multimodal-lora-training --skill h3-multimodal-lora-training-chengyansen-ai

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve? Training LoRA adapters for MiniMax H3 joint audio-video models risks silent failures: reference media may never enter the conditioning chain, audio may be dropped from the loss, configs may be silently ignored, and the final checkpoint may be weaker than intermediate ones. This Skill turns those risks into a gated, hash-linked, fail-closed ten-step workflow with verifiable evidence at every stage. ## Core Features & Use Cases - Ten-step gated state machine: intake, rights/environment proof, route selection, dataset contract, audit, experiment lock, backend config resolution, smoke training, checkpoint selection, and inference validation with delivery packaging. - Three source-pinned backend adapters: generates executable configs for DiffSynth-Studio (true Ref2VA), AI Toolkit (low-memory consumer GPUs), and SimpleTuner (FL2VA/joint AV with official VRAM presets), verifying Git SHA and source file hashes before launch. - Dataset and rights auditing: validates JSONL manifests, rights ledgers, captions, leakage, fps/frame-grid/dimension rules, and 32 kHz stereo audio via ffprobe, refusing to train without training_allowed=true. - Use Case: A user with a licensed dataset of video clips and reference images wants a Ref2VA LoRA. The Skill initializes the project, audits the manifest, locks the experiment, generates a pinned DiffSynth config, runs an authorized smoke test, scores checkpoints, and validates 0/1/multi-reference inference samples before packaging the delivery manifest. ## Quick Start Ask the AI to use the h3-multimodal-lora-training skill to initialize an auditable H3 LoRA training project for your multimodal dataset and chosen route and backend.

Frequently Asked Questions about h3-multimodal-lora-training

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I train a LoRA for MiniMax H3 video and audio?▼

Initialize a project with scripts/h3_project.py init, build a JSONL manifest separating target and reference media, audit it with lint_captions.py and audit_dataset.py, lock the experiment, then generate a pinned backend config and run an authorized smoke test before full training.

Which backend should I use for H3 Ref2VA LoRA training?▼

DiffSynth-Studio is the production adapter for true Ref2VA with image, video, audio, and video-audio references. AI Toolkit supports image/video references on consumer GPUs but rejects standalone reference audio; SimpleTuner covers FL2VA and joint AV but does not claim Ref2VA.

Does a LoRA loading in a Ref2VA workflow prove reference conditioning was trained?▼

No. The Skill separates load_compatible, condition_trained, and audio_trained as distinct claims. Reference conditioning is only claimed when reference media verifiably entered the training conditioning sequence and zero/one/multi-reference and order-swap inference tests show interpretable differences.

What hardware and dependencies are required for H3 LoRA training?▼

The bundled scripts use only the Python standard library; training dependencies come from the chosen backend. ffprobe is required for media audit and final inference validation, and GPU training or inference requires an explicit authorization token before execution.

Why does the final checkpoint not get selected automatically?▼

Long training can strengthen the subject while degrading base motion, composition, or audio ability. The Skill scores multiple checkpoints on fixed held-out prompts and seeds via score_checkpoints.py and selects the one passing route scorecard and base-preservation floors, which may be an intermediate checkpoint.

Can I use rsLoRA, DoRA, or other new LoRA variants with H3?▼

Standard LoRA is the production default. Variants like rsLoRA, LoRA+, DoRA, PiSSA, and Min-SNR are research-only and must pass implementation, weight, single-variable, and inference round-trip gates in the selected backend before any expanded claim.

Related Skills

Explore More Software Engineering Skills→
calesthiocalesthio
123

media-provenance-rights

Document rights, consent, provenance, and release risks in AI-generated media.

Community
Advanced
corca-aicorca-ai
9

release

Automate version bumps, manifests, and operator update instructions across a repository.

Official
Advanced
AutoByteusAutoByteus
2

research-engineer

Runs the research loop from paper understanding through implementation, experiments, and analysis.

Official
Advanced
BagelHoleBagelHole
46

llm-fine-tuning

Set up QLoRA, LoRA, and full fine-tuning infrastructure with Hugging Face TRL and Axolotl.

Community
Advanced