Rapport d’audit du skill
distributed-llm-pretraining-torchtitan Rapport d’audit.
Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.
Trust Score OpenAgentSkill
Trust Score OpenAgentSkill
The Trust Score helps an agent decide whether a skill is safe enough to shortlist before installation.
Adoption GitHub
Validé100
13K stars GitHub
Activité stars/forks
Validé92
13K stars et 948 forks; l’activité des issues n’est pas disponible dans les métadonnées actuelles
Maintenance récente
Validé100
3 jours depuis le dernier push
Clarté de licence
Validé86
MIT
Complétude README/SKILL.md
Validé94
Les métadonnées incluent suffisamment de contexte d’usage et de workflow
Risque dépendances/runtime
Avertissement44
command execution surface, credential or environment access
Disponibilité de l’installation
Validé92
npx skills add Orchestra-Research/AI-Research-SKILLs --skill distributed-llm-pretraining-torchtitan
Sécurité de la commande d’installation
Validé92
Chemin d’installation standard de package ou runtime
Surface de permissions
Échoué36
secrets or environment access, shell or command execution
Preuve du dépôt
Validé86
https://github.com/Orchestra-Research/AI-Research-SKILLs/tree/main/01-model-architecture/torchtitan
État de revue
Avertissement50.9
L’approbation de revue IA est absente
Résultats prouvés par Agent
Info54
Pas encore de données de résultats Agent
Vérifications
Revue d’installation et d’adoption
Chemin d’installation
92
npx skills add Orchestra-Research/AI-Research-SKILLs --skill distributed-llm-pretraining-torchtitan
Dépôt
88
https://github.com/Orchestra-Research/AI-Research-SKILLs/tree/main/01-model-architecture/torchtitan
Licence
86
MIT
Maintenance
100
3 jours depuis le dernier push
Revue IA
55
Review approval is missing
Complétude README/SKILL.md
94
Usable description available
Risque de dépendances
44
command execution surface, credential or environment access
Sécurité de la commande d’installation
92
Chemin d’installation standard de package ou runtime
Surface de permissions
36
secrets or environment access, shell or command execution
Activité stars/forks
92
13K stars et 948 forks; l’activité des issues n’est pas disponible dans les métadonnées actuelles
Adoption
88
13K stars GitHub
Avertissements
- Dependency or permission surface needs review
- Permission surface may require sandboxing
- L’approbation de revue IA est absente
- Quality score needs review
- Permission surface needs review: secrets or environment access, shell or command execution
- Dependency/runtime risk: command execution surface, credential or environment access
- Permission surface: secrets or environment access, shell or command execution
- Review status: AI review approval is missing
Méthode
This report combines public metadata, AI review output, repository freshness, install readiness, OpenAgentSkill events, quality scoring, trust checks, and the agent safety gate. It is not a full source-code security review.
Comparer les options proches
Skills associés à auditer ensuite
full-output-enforcement
Overrides default LLM truncation behavior. Enforces complete code generation, bans placeholder patterns, and handles token-limit splits cleanly. Apply to any task requiring exhaustive, unabridged output.
96K Stars · Rapport d’audit
peft-fine-tuning
Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. Use when fine-tuning large models (7B-70B) with limited GPU memory, when you need to train <1% of parameters with minimal accuracy loss, or for multi-adapter serving. HuggingFace's official library integrated with transformers ecosystem.
13K Stars · Rapport d’audit
litert-lm
Creates an Android app that runs an open text LLM on the device, on the CPU or the GPU, with the LiteRT-LM Kotlin API. Use this skill to build a new chat, summarization or extraction app with a .litertlm model from Hugging Face litert-community (Gemma, Qwen, Llama, Phi) - the dependency and manifest entries, getting the model file onto the device, engine initialization, a streamed multi-turn conversation, the ViewModel and screen.
459 Stars · Rapport d’audit