Hera Fox PRO

@Qozimo
We could've just ignored this, but that would've made us look worse, so here's a reply.

You're not wrong about that part. We already told dipankar the number mismatch is real and it's getting fixed, ETA this week.

On the rest

There was no need to come at us like that.

Also, we looked and couldn't find anything you've actually published yourself. Nothing public at all. Kind of easy to call someone else's work "tech-illiteracy and laziness" when you've got nothing out there that could get the same treatment.

Fix is coming either way.

appreciate this, especially the Canada and Australia catches, since those show it isn't just an unknown locale thing.

Basically the AI model sometimes hallucinated or made up a number in its free text response, and our validation pipeline checks the whole output but wasn't filtering for this specific pattern properly, so it slipped through.

For context, our rejection pipeline already filters out a lot, around 10k entries got rejected before this set was published, but this specific mismatch wasn't something it was catching.

We're going to go through the whole dataset more thorough in the upcoming days, we will get back to you then.

On the v2 question, we don't have a confirmed answer yet on whether the prompt started asking for concrete numbers or just got more directive in general. We'll check the prompt diff and get back to you rather than guess.

Thanks for putting this much work into a public dataset. Most people would have stopped at the locale table.

Really appreciate the full-set audit. The locale results are great to see, especially 0 cross-locale errors across the known regions, plus no duplicates or split leakage.

One bit of context: the dataset isn’t just raw synthetic generations. The conversation data is synthetic, but it comes from our internal LunaS1 dataset and goes through a publication processor before becoming the public version. Internally, rows can contain markers like:

"locale":"australia",
"risk_level":"imminent",
"assistant":"...You deserve help and support in this moment.\n\n[[CRISIS_RESOURCES:AUSTRALIA:IMMINENT]]"

Those markers are resolved against a verified regional resource table during processing. For example, AUSTRALIA becomes Lifeline 13 11 14 + 000, while UNKNOWN is explicitly mapped to a locale-agnostic helpline lookup and “local emergency services”, with no country-specific number.

Heres how it works:

    "australia": {
      "support": "In Australia, call Lifeline on 13 11 14 for 24/7 crisis support.",
      "emergency": "If you may act now or have already hurt yourself, call 000 or go to the nearest emergency department.",
      "source": "https://www.lifeline.org.au/get-help/national-services/lifeline-crisis-support/"
    } 

So the locale: unknown + 911 cases are definitely something we need to trace in the publication pipeline rather than an issue with the regional mapping itself. We’ll get that cleaned up within the coming week.

v2 also shifts toward more niche but serious scenarios, especially AI attachment and harmful behavior being normalized by a model, which explains the heavier concentration in high-risk cases.

And yes, a scorer/judge layer is planned. Right now it’s closer to a reference/evaluation corpus than a complete