Qwen3.5-4B — Reasoning Fine-Tune
A lightweight fine-tune of Qwen3.5-4B focused on improving reasoning, attention to detail, and epistemic behavior.
Clarity is the technical name I use for this personal experimental fine-tune. It is not intended to imply a formal benchmark result, standardized capability, or claim of superior model quality.
What changed
- Improved logical reasoning on everyday and structured problems
- Better multi-step problem solving and consistency checking
- Improved attention to contradictions and incorrect assumptions
- Some improvement in Python code precision
- Additional reasoning examples in English and Russian
The main training data was distilled from Qwen3.8-27B, combined with additional reasoning examples in English and Russian.
This is intended as a relatively subtle behavioral fine-tune rather than a major capability expansion. The goal was to make the model more deliberate and reliable when solving problems, while retaining the general capabilities of the original Qwen3.5-4B.
Note: This is a personal experimental fine-tune. The improvements described above are based primarily on my own qualitative evaluations rather than a standardized benchmark.
Evaluation
Initial qualitative testing suggests improvements in:
- constraint-based reasoning
- causal reasoning
- multi-step arithmetic
- error detection and correction
- distinguishing facts from assumptions
- avoiding invalid logical implications
The model can still produce incorrect reasoning and should not be treated as reliably calibrated. Some tendency toward overconfident or overly elaborate explanations was observed during evaluation.
Sampling parameters
General
- Temperature:
0.8 - Top-p:
0.95 - Top-k:
20 - Min-p:
0.05 - Repetition penalty:
1.05 - Presence penalty:
1.0
Code
- Temperature:
0.6 - Top-p:
0.95 - Top-k:
20 - Min-p:
0.0 - Repetition penalty:
1.00 - Presence penalty:
0.0
tags:
- gguf
- llama.cpp
- llama-cpp
- unsloth
Qwen3.5-4B-Clarity-1.0 : GGUF
This model was converted to GGUF format using Unsloth.
Example usage:
- For text only LLMs:
llama-cli -hf Regic1d3/Qwen3.5-4B-Clarity-1.0 --jinja - For multimodal models:
llama-mtmd-cli -hf Regic1d3/Qwen3.5-4B-Clarity-1.0 --jinja
Available model files:
Qwen3.5-4B-Distill-stage3-Inference.Q5_K_M.ggufQwen3.5-4B-Distill-stage3-Inference.Q8_0.gguf