---

name: data-and-configs description: "Guides agents to write, inspect, and debug Axolotl YAML configs, dataset formats, chat_template mappings, preprocessing, and config-schema issues." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Data and Configs

Use this sub-skill when an agent needs to write an Axolotl config YAML, choose or fix a dataset type, debug chat_template labels, inspect dataset columns, run axolotl preprocess --debug, or reason about axolotl config-schema output.

Fast Routing

• For YAML structure, required sections, preprocessing cache behavior, and config-schema workflows, use configuration.md.
• For datasets, pretraining_dataset, chat_template, alpaca, input_output, completion, pre-tokenized, and preference-data shape choices, use data-formats.md.
• For common errors and symptom-driven fixes, use troubleshooting.md.
• To preview a local JSON/JSONL chat dataset and infer field_messages plus message_property_mappings, run python scripts/inspect_chat_dataset.py data.jsonl using inspect_chat_dataset.py.
• To run a safe structural YAML check without importing Axolotl, run python scripts/validate_axolotl_config.py config.yaml using validate_axolotl_config.py.

Workflow

1. Classify the training/data family first: SFT-style datasets, streaming pretraining_dataset, preference data with rl, or fully pre-tokenized rows.
2. Map the raw columns to an Axolotl prompt strategy before editing hyperparameters; most data bugs come from type, field_messages, message_property_mappings, roles_to_train, or masking fields.
3. Use the bundled scripts for local structural checks only; they do not load models, tokenizers, remote datasets, or the Axolotl package.
4. Use axolotl config-schema or axolotl agent-docs for installed-package schema details when available.
5. Use axolotl preprocess config.yaml --debug when tokenized samples, labels, EOS/EOT handling, or sample packing must be confirmed with the real tokenizer and prompt strategy.

Boundaries

This sub-skill covers config shape, dataset loading fields, prompt strategy selection, preprocessing/debug handoffs, and schema-level validation. Route actual SFT/pretraining execution, preference-method tuning, GRPO/EBFT rewards, model/tokenizer/adapter quirks, distributed launch/performance, and broad CLI operations to the root skill router or the matching sibling sub-skill.