MMOPD/Qwen3-4B-OT3-tau2-3ep

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MMOPD/Qwen3-4B-OT3-tau2-3ep is a 4 billion parameter Qwen3-based language model developed by MMOPD, specifically fine-tuned as a tool-use domain teacher for customer service dialogues. It excels at agent-based interactions requiring tool calls and reasoning, particularly in airline, retail, and telecom scenarios. This model is the epoch-3 checkpoint from a supervised fine-tuning process on the inclusionAI/AReaL-tau2-data, demonstrating improved performance in tool-use benchmarks.

Loading preview...

Overview

MMOPD/Qwen3-4B-OT3-tau2-3ep is a 4 billion parameter Qwen3-based model, serving as the tool-use (τ²-bench) domain teacher within the MMOPD study. It is the third epoch checkpoint of a supervised fine-tuning (SFT) process, building upon MMOPD/Qwen3-4B-OT3-2ep.

Key Capabilities

  • Tool-Use Specialization: Fine-tuned on inclusionAI/AReaL-tau2-data, which comprises airline, retail, and telecom customer-service dialogues featuring tool calls and agent thinking.
  • Context Handling: Supports a 32,768-token context length, crucial for complex multi-turn dialogues.
  • Reasoning and Tool Calling: Trained to generate assistant turns that include both thinking processes (<think>...</think>) and structured tool calls (<tool_call>).
  • Performance on τ²-bench: Achieves a pass^1 score of 64.9% on the τ²-bench telecom base split using the official prompt, demonstrating its proficiency in agentic tasks.
  • Qwen3 Compatibility: Utilizes the native Qwen3 chat template and is designed to work with official τ²-bench tool schemas and vLLM's Qwen3 tool-call/reasoning parsers.

Training Details

The model was trained using SFT for 3 epochs on a filtered subset of inclusionAI/AReaL-tau2-data, specifically targeting the final assistant turn. Training involved a cosine learning rate schedule, bf16 precision, and ZeRO-2 optimization, with loss calculated only on the assistant's target output.

Good For

  • Agentic Applications: Ideal for developing AI agents that need to interact with users, understand context, reason, and make tool calls in customer service environments.
  • Customer Service Automation: Particularly suited for automating or assisting in customer support dialogues within telecom, airline, and retail sectors.
  • Research in Tool-Use LLMs: Provides a specialized model for researchers exploring tool-use capabilities and agent performance in complex conversational settings.

Finetunes
1 models