KANISHO
Community@hanasho744
KANISHO maintains 85 skills spanning LLM post-training and evaluation, academic research and paper-writing pipelines, scientific computing libraries, and business operations.
Agent Skills by KANISHO
Showing 75 vetted skills indexed across 1 GitHub repositories.
hanasho744openrlhf-training
Train large language models with distributed PPO, GRPO, RLOO, and DPO using Ray and vLLM.
hanasho744nemo-evaluator-sdk
Evaluates LLMs across 100+ benchmarks using containerized execution on Docker, Slurm, or cloud backends.
hanasho744slime-rl-training
Guides LLM post-training with reinforcement learning using Megatron-LM and SGLang.
hanasho744simpo-training
Train LLMs with reference-free SimPO preference optimization using alignment-handbook configs.
hanasho744obsidian-literature-workflow
Organizes project literature reviews in Obsidian from paper notes to synthesis and canvas graphs.
hanasho744distributed-llm-pretraining-torchtitan
Pretrains large language models at scale using TorchTitan with 4D parallelism on PyTorch.
hanasho744evaluating-llms-harness
Evaluates LLMs across 60+ academic benchmarks using standardized prompts and metrics.
hanasho744llama-factory
Guides fine-tuning of large language models using LLaMA-Factory with LoRA, QLoRA, and WebUI workflows.
hanasho744paper-self-review
Reviews academic papers for structure, claims, citations, and writing quality before submission.
hanasho744paper-claim-audit
Verifies every numeric claim in a research paper against raw result files using a zero-context cross-model reviewer.
hanasho744obsidian-source-ingestion
Ingests external materials into an Obsidian knowledge base as categorized source notes.
hanasho744implementing-llms-litgpt
Implements, fine-tunes, and deploys LLMs using LitGPT with LoRA, QLoRA, and FSDP training.
hanasho744fine-tuning-with-trl
Fine-tune language models with TRL using SFT, DPO, PPO, GRPO, and reward model training.
hanasho744evaluating-code-models
Benchmarks code generation models on HumanEval, MBPP, and MultiPL-E with pass@k metrics.
hanasho744citation-verification
Verifies academic citations against canonical scholarly sources like CrossRef, arXiv, and Semantic Scholar.
hanasho744verl-rl-training
Train LLMs with reinforcement learning using verl's HybridFlow framework.
hanasho744nature-citation
Segments manuscript text and finds Nature/CNS-family citations with RIS export.
hanasho744scientific-critical-thinking
Evaluate scientific claims, study methodology, and evidence quality using GRADE and Cochrane frameworks.
hanasho744grpo-rl-training
Implement GRPO reinforcement learning fine-tuning with TRL and custom reward functions.
hanasho744deep-research
Orchestrates a 13-agent pipeline for systematic academic research and APA 7.0 report generation.
hanasho744miles-rl-training
Configures RL training for large MoE models with FP8, INT4, and speculative decoding.
hanasho744mamba-architecture
Implements and benchmarks Mamba state-space models for linear-complexity sequence modeling.
hanasho744academic-paper
Generates academic paper drafts through a 12-agent pipeline with citations, bilingual abstracts, and peer review.
hanasho744rwkv-architecture
Implements RWKV models for linear-time inference with constant memory and infinite context.
Frequently Asked Questions About KANISHO
FAQPage SchemaWhat tasks can I accomplish with KANISHO's skill collection?ā¼
You can fine-tune and RL-train large language models (LoRA, QLoRA, PPO, GRPO, DPO, SimPO), benchmark models on 100+ academic and code evaluations, run autonomous research pipelines from idea discovery to paper PDF, verify citations, and perform scientific computing with RDKit, Biopython, SymPy, and Astropy.
Who is the target audience for these skills?ā¼
ML engineers post-training 7B-70B models on distributed GPU clusters, academic researchers writing and reviewing papers in English, Chinese, Japanese, Korean, or Spanish, and computational scientists working in cheminformatics, bioinformatics, materials science, astronomy, and applied microeconomics.
What are the installation and runtime requirements?ā¼
Training skills require PyTorch 2.0+, Transformers, and backends like Ray, vLLM, DeepSpeed, or Megatron-LM with multi-GPU clusters. Evaluation skills need Docker or Slurm. Scientific skills need Python 3.10+ with domain libraries; some NCBI Entrez and OpenRouter features require API keys and network access.
Are these skills open source and what do they cost?ā¼
Most skills carry MIT, BSD-3-Clause, or Apache-2.0 licenses and are free to use. Exceptions include the xlsx and pdf skills, which are proprietary, and MATLAB, which requires a MathWorks license though GNU Octave works as a free alternative.
Which frameworks do the LLM training skills depend on?ā¼
Skills build on OpenRLHF, TRL, LLaMA-Factory, Axolotl, Unsloth, PEFT, LitGPT, torchtitan, verl, slime, and torchforge. Common dependencies include torch, transformers, datasets, peft, accelerate, vllm, and ray, supporting Llama, Qwen, Gemma, Mistral, Phi, and DeepSeek model families.