How to use from
SGLang
Install from pip and serve model
# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Integral-Inference/integral-muse-copywriter" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Integral-Inference/integral-muse-copywriter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'
Use Docker images
docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Integral-Inference/integral-muse-copywriter" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Integral-Inference/integral-muse-copywriter",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'
Quick Links

integral-muse-copywriter

Deep-tech companies depend on documents that turn technical work into a case a customer or an investor can follow. The facts come from engineers, scientists and finance teams. The copy has to read as if a skilled writer made it, and every claim in it has to trace back to those facts.

General-purpose models write this kind of copy with habits that readers have learned to spot, such as inflated vocabulary and stock sentence shapes, and they often add claims the source never made. Readers discount the message once they notice. This model is a fine-tune of Muse-Glimmer-30B that writes from the facts it is given, in the plain and specific style of professional human copy.

How it was trained

This is Muse-Glimmer-30B with a LoRA of rank 64 on every linear layer of the text decoder, trained in two stages and merged into the weights. Supervised fine-tuning was tried first and thrown out. Professional copy carries facts its writers knew that the source they were handed does not contain, so imitating it taught the model to make things up, and the writing did not get better.

Stage one is ORPO on 541 pairs. The human copy uses only facts from its source: headlines, bill summaries, slogans, and sentences an editor cut puffery or jargon from. The rejected side is the base model's attempt.

Stage two is GRPO on 7,108 source documents, each with copy a professional wrote from it. The model writes eight drafts per document. A draft scores zero if a judge finds a claim the source does not support or if it drops more than a tenth of the supplied facts. The rest are scored on how their writing compares with the professional's, on hitting the requested length, and on staying inside the range of sentence shapes and stock phrases real copywriters use. In this stage the professional copy is never a training target. The model learns only from its own drafts.

Data

Every input is a document a person was handed and every reference is the copy a person wrote from it. The inputs and references come from 15 sources published before December 2022. No language model wrote any of it. Each input carries a one-line instruction that names the deliverable and the reader, states a target length and asks for the facts given and nothing more.

The repository carries pointers to the data, so anyone with the upstream data can rebuild the set and check it row by row. The full write-up is in docs/writeup.pdf, and the data source by source in docs/corpora.pdf.

File Contents
data/sources.tsv Each of the 15 sources: what the pair is, upstream URL, revision, license, rows per set
data/rl_prompts.tsv The 7,108 GRPO prompts: id, source, date, license, SHA-256 of the input and of the human copy
data/orpo_pairs.tsv The 541 ORPO pairs, same columns
data/dev.tsv The 253 dev pairs, same columns

Use

vllm serve Integral-inference/integral-muse-copywriter \
  --reasoning-parser muse_glimmer --max-model-len 8192
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
reply = client.chat.completions.create(
    model="Integral-inference/integral-muse-copywriter",
    temperature=0.7,
    extra_body={"chat_template_kwargs": {"reasoning_strength": "low"}},
    messages=[{"role": "user", "content":
        "Write a news release of roughly 300 words for general readers about this paper, "
        "based only on its text.\n\n" + paper_text}],
)
print(reply.choices[0].message.content)

Give it the deliverable, the reader, a target length and the source text, the way it was trained.

Downloads last month
-
Safetensors
Model size
30B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Integral-Inference/integral-muse-copywriter

Finetuned
(72)
this model