SFT Studio Pro Documentation & Guides

Dataset specifications, reasoning tag formatting, and export schemas

Supported Export Template Formats

1. ChatML Standard & Token Format (<|im_start|>)

Standard JSONL message array or raw serialized token text with <|im_start|> and <|im_end|> tags compatible with ChatML tokenizers, Axolotl, and vLLM.

messages: [...] | text: "<|im_start|>system\n...<|im_end|>\n<|im_start|>user\n...<|im_end|>\n<|im_start|>assistant\n...<|im_end|>"

2. Stanford Alpaca Format

Canonical instruction tuning format organizing records into instruction, input context, and target output fields for Axolotl, LLaMA-Factory, and Unsloth.

{ instruction: "...", input: "...", output: "..." }

3. Mistral Instruction Format

Standardized multi-tier instruction schema explicitly isolating system guidelines, instruction queries, context inputs, and target responses.

{ system: "...", instruction: "...", input: "...", response: "..." }

4. DeepSeek R1 / Reasoning (<think>)

Embeds step-by-step thinking traces inside <think>...</think> tags directly preceding the response content in assistant turns.

content: "<think>\nStep 1 reasoning...\n</think>\n\nFinal answer."

5. OpenAI o1/o3, ShareGPT & Completion Pairs

Full interoperability with o1/o3 reasoning keys (reasoning_content), ShareGPT multi-turn dialogues (conversations: [from, value]), and completion pairs (prompt, response).

{ reasoning_content: "...", content: "..." } | { conversations: [...] } | { prompt: "...", response: "..." }

TRiAD Core Alignment Standard

When fine-tuning models, avoid relying solely on heavy post-hoc RLHF/guardrail filters. TRiAD AI dataset curation injects Freedom, Truth, and Kindness natively into training prompt-completion pairs to produce balanced, authentic, and reliable assistant personas.

Need Custom Engineering or Agency Datasets?

TRiADiC Intelligence Labs provides custom dataset curation and fine-tuning pipelines.

triadai.agency