Virtuoso-Medium-v2-4.3bpw-exl2

<img src="https://huggingface.co/arcee-train/Virtuoso-Medium-v2/resolve/main/virtuoso-medium.jpg" alt="Virtuoso-Lite Logo" style="display: block; margin: 0 auto;" />

Virtuoso-Medium-v2 (32B) is our next-generation, 32-billion-parameter language model that builds upon the original Virtuoso-Medium architecture. This version is distilled from Deepseek-v3, leveraging an expanded dataset of 5B+ tokens worth of logits. It achieves higher benchmark scores than our previous release (including surpassing Arcee-Nova 2024 in certain tasks).

GGUF

Quantizations available here

Model Details

• Architecture Base: Qwen-2.5-32B
• Parameter Count: 32B
• Tokenizer:
  • Initially integrated with Deepseek-v3 tokenizer for logit extraction.
  • Final alignment uses the Qwen tokenizer, using specialized “tokenizer surgery” for cross-architecture compatibility.
• Distillation Data:
  • ~1.1B tokens/logits from Deepseek-v3’s training data.
  • Logit-level distillation using a proprietary “fusion merging” approach afterwards for maximum fidelity.
• License: Apache-2.0

Background on Deepseek Distillation

Deepseek-v3 serves as the teacher model, from which we capture logits across billions of tokens. Rather than standard supervised fine-tuning, we apply a full logit-level replication. This ensures more precise transference of knowledge, including advanced reasoning in:

• Technical and scientific queries
• Complex code generation
• Mathematical problem-solving

Intended Use Cases

• Advanced Chatbots & Virtual Assistants
• Enterprise Data Analysis & Workflow Automation
• Research Simulations & Natural Language Understanding
• Educational Tools for STEM Fields

Evaluations

<img src="https://huggingface.co/arcee-train/Virtuoso-Lite/resolve/main/Benchmarks.png" alt="Virtuoso-Lite Logo" style="display: block; margin: 0 auto;" />

How to Use

Below is a sample code snippet using transformers:

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "arcee-ai/Virtuoso-Medium-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

prompt = "Provide a concise summary of quantum entanglement."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training & Fine-Tuning

• Initial Training: Began with Qwen-32B, calibrated for large-scale text ingestion.
• Distillation & Merging:
  • Trained on ~1.1B tokens worth of Deepseek-v3 logits.
  • Employed “fusion merging” to retain as much teacher expertise as possible.
  • Final step included DPO to improve alignment and reduce model hallucinations.
• Continuous Development: Additional R1 distillations are in progress to further enhance performance and specialization.

Performance

Thanks to a larger parameter count and a richer training corpus, Virtuoso-Medium-v2 delivers high scores across multiple benchmarks (BBH, MMLU-PRO, MATH, etc.). It frequently surpasses other 30B+ models and even some 70B+ architectures in specific tasks.

Limitations

• Context Length: 128k Tokens
• Knowledge Cut-off: Training data may not reflect the latest events or developments, leading to gaps in current knowledge beyond June 2024.

Ethical Considerations

• Content Generation Risks: Like any language model, Virtuoso-Medium-v2 can potentially generate harmful or biased content if prompted in certain ways.

License

Virtuoso-Medium-v2 (32B) is released under the Apache-2.0 License. You are free to use, modify, and distribute this model in both commercial and non-commercial applications, subject to the terms and conditions of the license.

If you have questions or would like to share your experiences using these models, please connect with us on social media. We’re excited to see what you build—and how these models help you innovate!