model_bluepulse

A domain-aligned adapter for Qwen/Qwen2.5-1.5B-Instruct, built with the Nugen API.

model_bluepulse was aligned to its domain from its owner's own data using Nugen's train-time alignment. Served through the Nugen API, the one model id answers chat, embeddings and reranking, and every chat answer carries a confidence score that flags when a question falls outside the aligned domain.

Model details

• Developed by: [More Information Needed]
• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Nugen model id: model_01m6xcgy6n0qk91x
• Aligned by: Nugen Intelligence with train-time and inference-time alignment

This repository holds only the adapter weights from train-time alignment. The base model weights are not included. Inference-time alignment, confidence scores, embeddings and reranking run on the Nugen API.

Quick start

Get an API key from platform.nugen.in under API Keys. The model must be deployed and visible to your organisation. The same model value works for all three calls.

import requests

API_KEY = "<YOUR NUGEN API KEY>"
MODEL = "model_01m6xcgy6n0qk91x"
BASE = "https://api.nugen.in/api/v3/inference"
headers = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

Chat completions

r = requests.post(f"{BASE}/chat/completions", headers=headers, json={
    "model": MODEL,
    "messages": [{"role": "user", "content": "Summarise the key findings from the latest report."}],
    "max_tokens": 200,
    "temperature": 0.7,
    "stream": False,
})
resp = r.json()
print(resp["choices"][0]["message"]["content"], resp["confidence_score"])

Embeddings

r = requests.post(f"{BASE}/embeddings", headers=headers, json={
    "model": MODEL,
    "input": "performance benchmarks for the updated processing pipeline",
})
print(r.json()["data"][0]["embedding"][:5])

Reranking

r = requests.post(f"{BASE}/reranker", headers=headers, json={
    "model": MODEL,
    "query": "What caused the improvement in processing latency?",
    "documents": [
        "Batch sizes were reduced, cutting average latency by 35%.",
        "The team migrated to a new storage backend last quarter.",
        "Three additional engineers joined the platform team in June.",
    ],
    "top_n": 2,
})
print(r.json())

Why one model for all three

Most teams train and run a separate generation model, a separate embedding model and a separate reranker. Nugen aligns the model's internal representations to your domain in one training pass, so the same domain knowledge that improves generation also shapes its embeddings and reranking. One model id, one deployment, nothing extra to train or maintain. At serving time, inference-time alignment steers each answer and scores its confidence.

What the Developer edition includes

• Align a model from your own documents in minutes.
• One model id for chat, embeddings and reranking.
• A confidence score on every answer.
• Generated training data, a benchmark and an evaluation for each model.
• Deploy when you need it and undeploy when you do not.
• Export the adapter to your own Hugging Face account, like this one.

Your data stays yours. Models run on dedicated capacity with zero data retention by default.

Align your own model →

About Nugen Intelligence

Nugen Intelligence is a Mumbai and San Francisco based AI research company, backed by the Antler AI Disrupt inaugural fund, with a founding team from Stanford, Mines ParisTech, IIT Bombay and University of Pune. Our research in train-time representation engineering, inference-time steering and unified multi-task serving makes models behave with precision and consistency in high-stakes domains.

Licence and attribution

This adapter is a derivative work of Qwen/Qwen2.5-1.5B-Instruct. The base model's licence terms apply to its use and redistribution. See the base model card for them.

Learn more

• Documentation: docs.nugen.in
• Platform and API keys: platform.nugen.in
• Sign up: nugen.in/signup
• Enterprise: teams with larger or more specialised needs can write to support@nugen.in