drstupidity/granite-3.3-2b-customer-support

TEXT GENERATIONPricing:Input $0.32 / Cached $0.016 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kPublished:Sep 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The drstupidity/granite-3.3-2b-customer-support model is a 2 billion parameter language model, fine-tuned from ibm-granite/granite-3.3-2b-instruct. It is specifically optimized for customer support interactions, trained on the Bitext customer-support dataset to provide helpful and professional responses. This model excels at accurately addressing customer requests while maintaining an empathetic tone and avoiding the invention of details. Its primary use case is as a specialized customer support assistant for single-turn inquiries.

Loading preview...

Overview

The drstupidity/granite-3.3-2b-customer-support model is a specialized 2 billion parameter language model, fine-tuned from the ibm-granite/granite-3.3-2b-instruct base. Its core purpose is to act as a customer support assistant, trained extensively on the Bitext customer-support dataset. The model is designed to understand customer needs and respond accurately, helpfully, and professionally, emphasizing empathy and concrete next steps without fabricating information.

Key Capabilities and Features

  • Customer Support Specialization: Fine-tuned specifically for handling customer service requests.
  • Professional and Empathetic Tone: Engineered to communicate in a helpful, professional, and empathetic manner.
  • Fact-Oriented Responses: Trained to avoid inventing account details, order numbers, or policies not provided in the prompt.
  • Robust Prompt Handling: Features a 15% system-prompt dropout during training, allowing it to behave sensibly even with varied or absent system prompts.
  • Refusal of Off-Topic Requests: Includes synthetic 'decline' examples in its training to improve its ability to refuse out-of-domain or inappropriate requests, enhancing safety and reliability.

Training Details

The model was trained using LoRA (r=16, alpha=32, dropout=0.05) targeting key attention and feed-forward modules. It utilized a cosine learning rate schedule with a warmup, an effective batch size of 32, and a maximum sequence length of 640. Training data included 31,343 customer support rows and 1,235 synthetic 'when not to help' examples, crucial for teaching the model appropriate refusal behaviors. The data split was performed by near-duplicate clusters to prevent data leakage.

Limitations

  • Trained exclusively on synthetic, English, single-turn support data; it lacks multi-turn conversational training.
  • Does not have access to real account systems.
  • While trained not to invent facts, this is a tendency, not a guarantee.
  • Non-English input is outside its training distribution.