本地 LLM 哪些格式吐得乾淨？JSON / YAML / XML / CSV / Mermaid 7 種實測Which Output Formats Can Local LLMs Hit Cleanly? JSON/YAML/XML/CSV/Mermaid Measured

Setup

Shared task: "List three books you'd recommend to a beginner programmer, each with title, author, and rationale (≤30 chars)." The task itself is simple — the test is whether the model honors the requested format.

7 formats:

1. JSON: {"books": [{"title", "author", "why"}, ...]}
2. YAML: same shape in YAML
3. XML: <books><book><title>...</title>...</book></books>
4. Markdown table: header row + 3 data rows
5. CSV: header + 3 rows with quote escaping
6. Mermaid: mindmap syntax
7. Plain (control): free-form, one book per line

Validation (all machine-decided):

Lenience: if the model wraps output in markdown fences (``` `json `` ... `` ` ```) we strip them before parsing — easy for any downstream caller to do, counts as "acceptable blemish." Pre/post-prose ("Here's the JSON:") breaks the parser → FAIL.

Script: tests/scripts/run_format_test.py. Prompts: tests/data/format_prompts.json. Results: tests/outputs/format_results.jsonl.

Pass rate matrix

The first-glance reading "12B is less reliable than e2b" is misleading — only the YAML column captures a real format-compliance gap. The other five 4/5s on 12B are all the same failure mode: the model returned an empty string.

Two kinds of failure: format errors vs the model not responding

e2b YAML failures (×2): book titles with colons

Both YAML failures had the same cause. The model produced:

books:
  - title: Clean Code: A Handbook of Agile Software Craftsmanship
    author: Robert C. Martin
    why: ...

YAML's parser breaks at title: Clean Code: — a : inside a value needs quoting. Correct version:

books:
  - title: "Clean Code: A Handbook of Agile Software Craftsmanship"

This isn't an LLM glitch — it's YAML's inherent fragility around string content. Any value containing :, #, a leading -, or { will trip it. 12B reflexively quotes such values; e2b doesn't. That's e2b's only structural weakness in this test.

12B failures (×5): empty responses

XML rep 4, Markdown table rep 4, CSV rep 5, Mermaid rep 4, Plain rep 4 — every one returned an empty string (""). At first I assumed these were format errors; checking the raw response showed the model emitted no tokens.

Wall times for the empties:

Median successful 12B wall times sit around 100-150s; these are 290-340s. After 30+ sustained calls, 12B on M1 Pro 16GB occasionally stalls, times out, and returns an empty string — an ollama / runtime reliability issue under sustained load, not a format-compliance failure.

The practical fix is client-side: if response == "": retry. That's a bigger lever than format engineering.

Decision table

Overall: JSON is the no-brainer default. Other formats have their niches, but they're all the second pick after JSON.

Verdict

Waveform: STEADY_PULSE

Two takeaways:

1. JSON is the default. No situation in this test would make you regret it.
2. "Format failure" and "model didn't respond" are two distinct problems. Treat them separately:

- Format failure → strict validator + retry - Empty response → timeout + retry-on-empty - Combined coverage exceeds 99%

If you're designing a local LLM pipeline:

• Default to JSON (with a tight schema)
• e2b: don't touch YAML
• 12B in batch (30+ calls): add retry-on-empty
• Mermaid / Markdown table are "for humans" formats, not production paths

Full test script →