ahmetggg/Luck-Qwen3-4b-Code-FineTune

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The ahmetggg/Luck-Qwen3-4b-Code-FineTune is a 4.05 billion parameter QLoRA fine-tune of unsloth/Qwen3-4B-Instruct-2507, specialized in agentic coding, multi-step reasoning, and tool use. This model enhances output formatting, tool interaction, and calibrated uncertainty, allowing it to refuse answers when knowledge is lacking. It is optimized for tasks requiring repository-level code navigation, issue resolution, and structured function calling within a 4,096 token context length.

Loading preview...

Model Overview

ahmetggg/Luck-Qwen3-4b-Code-FineTune is a 4.05 billion parameter model, fine-tuned using QLoRA (4-bit NF4) on the unsloth/Qwen3-4B-Instruct-2507 base. It focuses on improving specific aspects of model behavior rather than increasing raw capability. The fine-tuning process involved 1,267 steps over one epoch, resulting in a significant reduction in training loss from ~1.7 to ~0.55.

Key Capabilities

  • Agentic Coding: Enhanced for repository-level navigation, issue resolution, and code generation.
  • Multi-Step Reasoning: Improved performance in step-by-step problem-solving across science and mathematics domains.
  • Tool Use: Specialized in structured, multi-turn function and tool calling interactions.
  • Calibrated Uncertainty: Designed to refuse to answer ("I don't know") when encountering genuine knowledge gaps, promoting more reliable responses.

Training Details

The model was trained on a diverse dataset including:

  • nvidia/Open-SWE-Traces: Filtered for resolved agentic coding trajectories.
  • open-thoughts/OpenThoughts3-1.2M: Focused on math and science reasoning.
  • Agent-Ark/Toucan-1.5M: Provided real multi-turn tool interaction trajectories.
  • Custom R-Tuning Set: Used to calibrate uncertainty by relabeling incorrect MMLU outputs as "I don't know" responses.

Good For

This model is particularly well-suited for applications requiring:

  • Consistent output formatting in code-related tasks.
  • Reliable interaction with external tools and functions.
  • Intelligent refusal to answer when information is unavailable, reducing hallucination.
  • Complex problem-solving that benefits from structured reasoning.

Finetunes
2 models