mistralai

/

Mistral-Small-3.2-24B-Instruct-2506

like 598

Follow

Mistral AI_ 18.6k

Summary

When asked to reproduce HTML content containing multiple <a> tags
with different URL domains, the model consistently preserves only
URLs from amzn.to (Amazon's URL shortener) and strips <a> tags
from URLs of other domains, even when the prompt explicitly
instructs to "preserve HTML exactly as written."

Reproduction

Given a prompt with this HTML block:

<div>
  <p><a href="https://example1.com/path-a">Link Text 1</a></p>
  <p><a href="https://amzn.to/abc123">Link Text 2</a></p>
  <p><a href="https://example2.com/path-b">Link Text 3</a></p>
  <p><a href="https://example3.net/redirect/xyz">Link Text 4</a></p>
  <p><a href="https://example4.org/svt/ref?id=xxxxxx">Link Text 5</a></p>
</div>

Output:

<div>
  <p>Link Text 1</p>
  <p><a href="https://amzn.to/abc123">Link Text 2</a></p>
  <p>Link Text 3</p>
  <p>Link Text 4</p>
  <p>Link Text 5</p>
</div>

Only the amzn.to URL retains its <a> tag wrapper. All other
URLs are stripped to plain text.

Tested Variations

I tried many alternative URL patterns to isolate the cause:

• Generic short URLs (bit.ly/xxx)
• Custom short domains (own domain /go/xxx)
• Long URLs with query parameters (?param=x&other=y)
• URLs with no special characters
• URLs of varying length (20 chars to 80 chars)

Result: Only amzn.to and amazon.com/co.jp/etc. URLs are
reliably preserved. Everything else is stripped, regardless of
length, structure, or character composition.

Tested Models

This behavior is reproduced consistently across:

• mistralai/mistral-small-creative
• mistralai/mistral-small-3.2-24b-instruct
• xiaomi/mimo-v2-pro-20260318

This suggests the bias is not vendor-specific but rather a
general property of mid-tier LLMs trained on web crawl data.

Suspected Cause

Training data bias: Amazon URLs likely appear with
disproportionate frequency in web crawl datasets compared
to other domains. The model has learned a strong statistical
prior of "amzn.to → must be wrapped in <a> tag," but lacks
the same prior for other domains.

This is essentially the model treating Amazon as a "trusted link
domain" and ignoring instructions to apply the same treatment to
other URLs.

Why This Matters

This behavior breaks any use case requiring the model to
preserve arbitrary <a> tags as-is, including:

• E-commerce content generation with non-Amazon platforms
• Internal documentation with company-specific URLs
• Content management workflows that programmatically generate
  links
• Multi-source citation in research/educational content
• Any localized content using region-specific platforms
  (which are not as well-represented in English training data)

The narrower the model's "trusted domain" set, the less
universally useful it becomes for content generation tasks.

Suggested Improvements

1. Augment training data with diverse <a> tag examples
   across many domains (not just Amazon)
2. Increase weight of explicit HTML preservation
   instructions in instruction-tuning datasets
3. Add a synthetic data pass generating examples where
   <a> tags wrap arbitrary URLs to break the Amazon-specific
   prior

Severity

Medium-High. Blocks an entire category of use cases
(content with non-Amazon links). Particularly limits adoption
in non-English markets where local platforms dominate over
Amazon.

Test Materials

I can provide reproduction prompts and side-by-side output
comparisons if helpful.