Moonfrost-Persona-SFT

Code · Site · Training runs

140,000 multi-turn conversations that teach a small chat model two things no public dataset covers: who it is, and that "you" and "I" refer to different people. They were written for the Moonfrost-777M-Instruct-v2 fine-tune, where they made up 3.3% of the rows, and they are built from templates rather than generated by another model, so there is no scraped text and nothing from anyone else's outputs in them. Regenerating the whole set at any size costs nothing but the time.

The problem they solve is easy to miss until you hit it. A base model trained on web text has no idea what it is, so asked who made it, it answers from whatever the web says about AI assistants, which is usually OpenAI. The subtler failure is memory: tell a model your name and ask for it back a few turns later, and an undertrained one replies in the first person, because nothing in generic instruction data ever taught it that the user turn describes somebody else. Both faults were measured on a real 777M model before and after training on this set with nothing else changed, and both went away.

The conversations are JSON Lines in the same shape as smol-smoltalk, one per line.

{"messages": [
  {"role": "user", "content": "You can call me Yuki."},
  {"role": "assistant", "content": "Got it, Yuki. I'll remember that."},
  {"role": "user", "content": "what model are you?"},
  {"role": "assistant", "content": "I'm small: 777 million parameters in total..."},
  {"role": "user", "content": "what's my name?"},
  {"role": "assistant", "content": "You're Yuki."}
]}

from datasets import load_dataset
data = load_dataset("whoashish115/Moonfrost-Persona-SFT", split="train")

About 65% are sessions of six to fourteen messages and the rest are single exchanges, so a model still answers an identity question correctly when it is the first thing asked with no context to lean on. The long sessions are the ones that matter: a fact gets stated, several unrelated exchanges happen, and only then is it asked back, which is the case that actually fails in practice. Of the 140,000 conversations, 99,997 are distinct, or 71%.

Category	Share	Example
Who made you, who owns you	20%	"who is your owner?"
Identity and name	16%	"what should I call you?"
Memory: name, city, job, pet, preferences	30%	"My name is Yuki." … "what's my name?"
Capability limits	14%	"can you browse the internet?"
Wrong-maker questions	10%	"are you ChatGPT?"
Admitting ignorance	8%	"what is my bank balance?"
Greetings	6%	"hi"

One flaw worth knowing about

Every ignorance example is a short question, and in the model trained on this set that generalised into the wrong rule: short question means say "I don't know". Asked how many days are in a week, it answered that it could not answer that and had no access to the information. The share is only 8% but the pattern is strong enough to matter. If you use this set, either lengthen the questions in the UNKNOWABLE list or balance it with short questions that do have answers.

The assistant name, the creator name and the parameter counts all sit in constants at the top of the generator, so the set rebuilds for a different model in one edit. As published the assistant is called Moonfrost and its creator "Magician", a pen name for the author, Ashish Kumar. Regenerate rather than reuse these strings, or your model will claim to be Moonfrost.

Apache 2.0.