Docs
Sign inJoin the waitlist

Export your model

Everything Facet builds for you is yours. Export a Model to keep a copy, run it on your own hardware, or move it elsewhere. Exporting is free.

Export

  • In the web app: open the Model’s Settings, and under Export and archive choose Download. You get a zip of the Model’s live version.
  • From your agent: export_model returns the same Instructions, evals and examples as data. The trained files come only in the zip.

What’s inside

File What it is
README.md The base model and the exact revision (a fixed snapshot of it) this version was built on
adapter/ The part Facet trained for your Model: a LoRA adapter, the standard small add-on to a base model, in the PEFT format most tools read. Absent if the version only changed Instructions
instructions.json The system prompt, example answers and output format Facet sends with every call
training.jsonl The data Facet trained on, in OpenAI chat format, one example per line. Absent if the version only changed Instructions
evals.jsonl One line per example: its input, its ideal answer, and its eval: the statement, how it’s graded and its bar

Hidden examples are not included: they stay in Facet, so its evals stay honest.

Run it yourself

The adapter loads on its base model with any tool that reads PEFT adapters. With vLLM, an open source server, on a machine with a GPU:

vllm serve <base model from README.md> --revision <revision from README.md> \
  --enable-lora --lora-modules support-triage=./adapter

Then call it as support-triage with any OpenAI SDK, and send the system prompt from instructions.json as the first message: Facet adds it to every call for you, but a server you run won’t.

With Hugging Face Transformers:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

name, rev = "<base model from README.md>", "<revision from README.md>"
base = AutoModelForCausalLM.from_pretrained(name, revision=rev)
model = PeftModel.from_pretrained(base, "adapter")
tokenizer = AutoTokenizer.from_pretrained(name, revision=rev)

system = "<the system prompt from instructions.json>"
messages = [{"role": "system", "content": system}, {"role": "user", "content": "I was charged twice"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
print(tokenizer.decode(model.generate(inputs, max_new_tokens=200)[0][inputs.shape[1]:], skip_special_tokens=True))

Use the base model and revision named in README.md; an adapter only works on the base it was built on.

Next steps

Esc