跳到正文
cccskills
免费在 GitHub 公开

axolotl

Routes agents working with Axolotl config-driven LLM fine-tuning, datasets, SFT, preference tuning, GRPO/EBFT, model adapters, distributed training, CLI operations, and troubleshooting.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill
查看安装方法

包含的文件（44）

  • SKILL.md4.8 KB
  • references/repo-provenance.md1.9 KB
  • references/repo-routing-metadata.json566 B
  • references/troubleshooting.md3.7 KB
  • scripts/check_axolotl_environment.py3.9 KB
  • sub-skills/cli-and-operations/references/cli-reference.md5.2 KB
  • sub-skills/cli-and-operations/references/operations.md7.4 KB
  • sub-skills/cli-and-operations/references/troubleshooting.md7.2 KB
  • sub-skills/cli-and-operations/scripts/axolotl_command_builder.py6.7 KB
  • sub-skills/cli-and-operations/scripts/check_axolotl_install.py6.4 KB
  • sub-skills/cli-and-operations/SKILL.md2.9 KB
  • sub-skills/data-and-configs/references/configuration.md5.2 KB
  • sub-skills/data-and-configs/references/data-formats.md6.4 KB
  • sub-skills/data-and-configs/references/troubleshooting.md5.8 KB
  • sub-skills/data-and-configs/scripts/inspect_chat_dataset.py13.0 KB
  • sub-skills/data-and-configs/scripts/validate_axolotl_config.py19.0 KB
  • sub-skills/data-and-configs/SKILL.md2.6 KB
  • sub-skills/distributed-and-performance/references/distributed-training.md7.5 KB
  • sub-skills/distributed-and-performance/references/performance-options.md6.3 KB
  • sub-skills/distributed-and-performance/references/troubleshooting.md6.7 KB
  • sub-skills/distributed-and-performance/scripts/check_distributed_config.py19.3 KB
  • sub-skills/distributed-and-performance/SKILL.md4.0 KB
  • sub-skills/model-loading-and-adapters/references/adapters-and-quantization.md7.3 KB
  • sub-skills/model-loading-and-adapters/references/model-selection.md8.1 KB
  • sub-skills/model-loading-and-adapters/references/new-model-support.md7.7 KB
  • sub-skills/model-loading-and-adapters/references/troubleshooting.md6.9 KB
  • sub-skills/model-loading-and-adapters/scripts/check_model_config.py23.3 KB
  • sub-skills/model-loading-and-adapters/SKILL.md4.1 KB
  • sub-skills/preference-tuning/references/data-formats.md7.3 KB
  • sub-skills/preference-tuning/references/troubleshooting.md8.5 KB
  • sub-skills/preference-tuning/references/workflows.md7.9 KB
  • sub-skills/preference-tuning/scripts/check_preference_dataset.py12.3 KB
  • sub-skills/preference-tuning/SKILL.md3.4 KB
  • sub-skills/rl-and-rewards/references/ebft.md5.0 KB
  • sub-skills/rl-and-rewards/references/grpo.md6.2 KB
  • sub-skills/rl-and-rewards/references/reward-functions.md5.1 KB
  • sub-skills/rl-and-rewards/references/troubleshooting.md6.4 KB
  • sub-skills/rl-and-rewards/scripts/validate_reward_function.py9.4 KB
  • sub-skills/rl-and-rewards/SKILL.md3.5 KB
  • sub-skills/sft-and-pretraining/references/training-options.md6.0 KB
  • sub-skills/sft-and-pretraining/references/troubleshooting.md7.6 KB
  • sub-skills/sft-and-pretraining/references/workflows.md6.1 KB
  • sub-skills/sft-and-pretraining/scripts/make_minimal_sft_config.py8.3 KB
  • sub-skills/sft-and-pretraining/SKILL.md3.8 KB

SKILL.md（原文）

安装前可以查看交给 AI 代理的指令内容。

Axolotl

Use this repo skill when the user asks about Axolotl, axolotl CLI commands, Axolotl YAML configs, LLM fine-tuning recipes, dataset/preprocess issues, LoRA/QLoRA, preference tuning, GRPO/EBFT, vLLM-assisted training, model architecture support, DeepSpeed/FSDP/Ray scaling, or Axolotl-specific troubleshooting.

Axolotl is config-driven: a single YAML config selects the base model, adapter/training method, datasets, preprocessing, launch/runtime settings, and output paths.

Start Here

  1. Identify the user’s primary task family before editing YAML.
  2. Read the nearest sub-skill route below, then only open deeper references as needed.
  3. Prefer safe static helpers first; run axolotl preprocess <config> --debug only in the user’s Axolotl environment when tokenizer/model access and runtime cost are acceptable.
  4. Treat bundled helper scripts as triage tools, not replacements for Axolotl’s installed config-schema, preprocess, or training runtime.
  5. Read references/repo-provenance.md before judging whether this skill matches a current Axolotl checkout.

Route By Task

User taskRead
Write or debug YAML config structure, dataset columns, chat_template, prompt strategies, preprocessing, sample packing, or schema issuesdata-and-configs
Build SFT, LoRA/QLoRA, full fine-tune, or continual pretraining recipes; diagnose SFT/pretraining loss, OOM, checkpoint, or resume issuessft-and-pretraining
Configure DPO, IPO, KTO, ORPO, SimPO, reward model, process reward model, paired/unpaired preference data, or preference loss settingspreference-tuning
Configure GRPO, EBFT, custom reward functions, vLLM rollout servers, async online RL, NeMo Gym, or Hatchery reward hooksrl-and-rewards
Choose base models, tokenizers/processors, chat templates, LoRA targets, QLoRA, quantization/QAT/PTQ, multimodal settings, architecture quirks, or new model supportmodel-loading-and-adapters
Configure DeepSpeed, FSDP, Ray, SLURM/multi-node, tensor/context/sequence/expert parallelism, precision, kernels, profiling, or performance troubleshootingdistributed-and-performance
Use axolotl commands, launchers, agent-docs, config-schema, fetch, inference, evaluate, merge-lora, vllm-serve, quantize, install checks, or command constructioncli-and-operations

Common Entry Points

  • axolotl preprocess config.yaml --debug checks rendered/tokenized samples and label masking before expensive training.
  • axolotl train config.yaml launches training; use --launcher torchrun, --launcher accelerate, or --launcher python deliberately.
  • axolotl inference config.yaml, axolotl merge-lora config.yaml, axolotl evaluate config.yaml, axolotl vllm-serve config.yaml, and axolotl quantize config.yaml are operational routes covered by cli-and-operations.
  • axolotl agent-docs --list, axolotl agent-docs <topic>, and axolotl config-schema --field <field> are preferred installed-package truth checks when available.

Cross-Cutting Helpers

Routing Notes

  • Reward-model and process-reward-model training belong to preference-tuning; callable reward functions for GRPO/EBFT belong to rl-and-rewards.
  • Config validation usually starts in data-and-configs, then narrows to method/model/distributed/CLI-specific static helpers, then escalates to axolotl config-schema or axolotl preprocess in the user’s runtime.
  • Model downloads, gated repositories, GPU kernels, vLLM services, distributed launchers, and long training runs require user environment readiness and should not be inferred from static helper success.

Refresh Guidance

Run refresh-repo-skill when the Axolotl commit, package version, CLI entry points, config schema, docs, examples, or training APIs have changed relative to references/repo-provenance.md.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护，适用 Google 的隐私权政策和服务条款。

同一仓库的技能

3d-resnets-pytorch

免费

Routes 3D ResNets PyTorch video action-recognition workflows across training, inference, and data preparation.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

3ddfa

免费

Guide 3DDFA Python inference, geometry rendering, training/evaluation, and optional C++ ONNX workflows for 3D dense face alignment.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

3ddfa-v2

免费

Routes 3DDFA_V2 face-alignment setup, still-image demos, video tracking, and ONNX benchmarking workflows.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

ab3dmot

免费

Operate AB3DMOT 3D multi-object tracking workflows for KITTI and nuScenes data, tracking, evaluation, and visualization.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

accelerate

免费

Use Hugging Face Accelerate for PyTorch training-loop migration, distributed launch/configuration, DeepSpeed/FSDP/TPU backend setup, big-model inference/offload, checkpointing, tracking, and troubleshooting.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

acme

免费

Route Acme reinforcement-learning framework tasks across core loops, replay/data, JAX agents, and TensorFlow/Sonnet agents.

中文简介正在准备中。当前显示原文说明。

VectorSpaceLab/AREX-Skill3402026年9月3日 更新

查看 VectorSpaceLab 的全部技能

报告此技能的问题