wesjos/mario-html-sft

Mario HTML SFT (830 samples)

A small, cleaned SFT dataset for teaching a model to answer with HTML "cards" instead of Markdown. Each answer is an HTML fragment rooted at <div class="ai-view">, using a fixed component vocabulary (.card, .grid.grid-2, .callout, .stepper, .metric-group, ...). Styling is left to a front-end stylesheet, so the model only learns structure.

Companion tutorial: fine-tuning Qwen3.5-4B with Unsloth on a single 24GB GPU (see the GitHub repo nonwesjoe/SFT-LLM-mario-html).

File

File	Description
data/train.jsonl	830 records, one JSON object per line (JSON Lines)

Composition

	Count
Total samples	830
With chain-of-thought (has_cot = true)	730
Direct HTML, no thinking (has_cot = false)	100
Chinese instructions (lang = "zh")	130

The instructions are open-ended questions in the style of the LIMA dataset. Answers were generated by a larger teacher model and filtered (see below).

Schema

Each record has these fields:

Field	Type	Description
id	string	Unique id, e.g. lima_0000, zh-1029
index	int	Position index
instruction	string	The user question
has_cot	bool	Whether the sample has a reasoning trace
cot	string	Plain-text reasoning (empty when has_cot is false). Contains no HTML tags
html	string	The HTML card answer rooted at <div class="ai-view">
conversations	list	Two turns: {"from": "human", "value": ...} and {"from": "gpt", "value": ...}. The gpt value is <think>\n{cot}\n</think>\n{html} when has_cot is true, otherwise just html
lang	string	"zh" for Chinese samples; absent otherwise
cot_tokens, html_tokens, total_tokens	int	Token counts (present in 700 records)
finish_reason	string	Generation stop reason, e.g. stop (present in 700 records)
elapsed_s	float	Generation time in seconds (present in 700 records)
reasoning_effort	string	Teacher reasoning setting (present in 330 records)

Some metadata fields only appear in a subset of records, so do not assume they are present everywhere.

Example (abridged):

{
  "id": "lima_0000",
  "instruction": "Can brain cells move? By movement I mean long distance migration ...",
  "has_cot": true,
  "cot": "The user is asking about whether brain cells ... can migrate ...",
  "html": "<div class=\"ai-view\">\n\n<h2>Can Brain Cells Migrate Long Distances?</h2>\n...",
  "conversations": [
    {"from": "human", "value": "Can brain cells move? ..."},
    {"from": "gpt",   "value": "<think>\nThe user is asking ...\n</think>\n<div class=\"ai-view\">...</div>"}
  ]
}

Quality rules

Every sample passes these checks:

1. The cot field contains zero HTML tags, checked with an HTML parser rather than a regex.
2. The gpt reply has exactly one <think>...</think> block, immediately followed by <div class="ai-view">.
3. Samples without reasoning start directly with <div class="ai-view">, and every tag is properly paired and closed.

Ten samples are technical questions whose thinking region legitimately contains code syntax (for example <xs:element> or List<T>). These were reviewed by hand and kept, because removing the syntax would change the meaning of the question.

Why this matters: an earlier version of the data had about 29% of samples with HTML drafts inside <think>. Models trained on that learned to sketch layouts while reasoning, and the habit persisted through DPO. Clean data was the only fix.

Usage

from datasets import load_dataset

ds = load_dataset("wesjos/mario-html-sft", split="train")
sample = ds[0]

messages = [
    {"role": "user", "content": sample["instruction"]},
    {"role": "assistant", "content": sample["conversations"][1]["value"]},
]

The tutorial trains with no system prompt on purpose, so the model answers in card format unconditionally.

Limitations

• Small: 830 samples. Suitable for format teaching and small experiments, not general capability.
• Answers come from a single teacher model and reflect its style.
• The HTML vocabulary is specific to this project. A different front-end needs its own component set.

License

Apache-2.0. The base model you train on keeps its own license (for the tutorial, Qwen3.5-4B).