---
license: apache-2.0
library_name: gguf
pipeline_tag: text-generation
base_model:
- naksyu/yui-math-python-qwen35-4b-v0.5d-fft
- Qwen/Qwen3.5-4B
datasets:
- drlee1/deepseek-v4-distill-ko-1k
- Jongsim/claude-opus-4.6-reasoning-12k-ko-filtered-v2
language:
- ko
- en
tags:
- gguf
- qwen3.5
- qwen35
- q6_k
- llama.cpp
- lm-studio
- math
- python
- korean
- reasoning
model_name: yui-math-python-qwen35-4b-v0.5d-fft-GGUF
---

# yui-math-python-qwen35-4b-v0.5d-fft GGUF

GGUF export of [`naksyu/yui-math-python-qwen35-4b-v0.5d-fft`](https://huggingface.co/naksyu/yui-math-python-qwen35-4b-v0.5d-fft), a full-parameter SFT of `Qwen/Qwen3.5-4B` for Korean/English math reasoning, Python-assisted calculation habits, and compact practical answers.

## Recommended File

Use this file:

| File | Quantization | Size | Status |
| --- | --- | ---: | --- |
| `yui-math-python-qwen35-4b-v0.5d-fft-no-mtp-Q6_K.gguf` | Q6_K | 3.23 GiB | validated with `llama-server` |

This is the intended LM Studio / llama.cpp file for normal local inference.

## Compatibility

This model uses the `qwen35` GGUF architecture. Use a recent `llama.cpp` build or an LM Studio build with Qwen3.5 GGUF support.

Important: the source Transformers config contains `mtp_num_hidden_layers=1`, but the exported checkpoint used here does not contain separate `mtp.*` tensors. A plain GGUF conversion can produce metadata that expects a non-existent `blk.32.*` layer and then fail during load.

This GGUF was therefore converted with `--no-mtp`. The validated metadata is:

| Field | Value |
| --- | --- |
| `general.architecture` | `qwen35` |
| `general.file_type` | `18` / Q6_K |
| `qwen35.block_count` | `32` |
| `qwen35.context_length` | `262144` |
| `qwen35.attention.head_count` | `16` |
| `qwen35.attention.head_count_kv` | `4` |
| `qwen35.rope.dimension_sections` | `[11, 11, 10, 0]` |
| `qwen35.nextn_predict_layers` | absent |
| tensor count | `426` |

## llama.cpp Example

CPU load smoke:

```bash
llama-server \
  -m yui-math-python-qwen35-4b-v0.5d-fft-no-mtp-Q6_K.gguf \
  --host 127.0.0.1 \
  --port 18080 \
  -c 2048 \
  -ngl 0
```

Use a larger context only when you have enough RAM/VRAM. The model config advertises 262K context, but this GGUF was only smoke-tested locally at a small context for load compatibility.

## LM Studio

In LM Studio, import or place:

`yui-math-python-qwen35-4b-v0.5d-fft-no-mtp-Q6_K.gguf`

If loading fails, first check that the runtime supports `qwen35`. If the error mentions missing `blk.32.*` tensors, it is the MTP metadata mismatch; use the `no-mtp-Q6_K` file rather than a plain conversion.

## Source Model

The source model card is here:

- [`naksyu/yui-math-python-qwen35-4b-v0.5d-fft`](https://huggingface.co/naksyu/yui-math-python-qwen35-4b-v0.5d-fft)

Base model:

- [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B)

## Training Data Summary

The source checkpoint was trained on `qwen35_sft_v0_5d_system_mix_with_all_data_cutoff2048.jsonl`.

| Data bucket | Retained rows |
| --- | ---: |
| DeepSeek-derived Korean sample | 462 |
| Claude/trace Korean reasoning dataset | 11,973 |
| Local/Yui/Lime/generated data | 37,994 |
| Total | 50,429 |

Public data sources included:

- [`drlee1/deepseek-v4-distill-ko-1k`](https://huggingface.co/datasets/drlee1/deepseek-v4-distill-ko-1k)
- [`Jackrong/DeepSeek-V4-Distill-8000x`](https://huggingface.co/datasets/Jackrong/DeepSeek-V4-Distill-8000x)
- [`Jackrong/GLM-5.1-Reasoning-1M-Cleaned`](https://huggingface.co/datasets/Jackrong/GLM-5.1-Reasoning-1M-Cleaned)
- [`Jongsim/claude-opus-4.6-reasoning-12k-ko-filtered-v2`](https://huggingface.co/datasets/Jongsim/claude-opus-4.6-reasoning-12k-ko-filtered-v2)
- [`Roman1111111/claude-opus-4.6-10000x`](https://huggingface.co/datasets/Roman1111111/claude-opus-4.6-10000x)
- [`nohurry/Opus-4.6-Reasoning-3000x-filtered`](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered)
- [`TeichAI/claude-4.5-opus-high-reasoning-250x`](https://huggingface.co/datasets/TeichAI/claude-4.5-opus-high-reasoning-250x)

## Limitations

- Experimental local SFT conversion, not a benchmarked production release.
- Q6_K quantization changes numerical behavior from the bf16 source checkpoint.
- Requires current GGUF runtime support for Qwen3.5 / `qwen35`.
- The model can still make arithmetic, reasoning, and code mistakes. Verify important outputs externally.

## Training Data

This model was fine-tuned for experimental math/Python reasoning behavior.

The training mix includes:
- User-created Lime/Yui math-python SFT data
- Public Hugging Face datasets including:
  - drlee1/deepseek-v4-distill-ko-1k (MIT)
  - Jongsim/claude-opus-4.6-reasoning-12k-ko-filtered-v2 (Apache-2.0)

Some upstream samples are synthetic or translated model outputs. This release is intended as a research/experimental small-model fine-tune, not as a commercial substitute for the upstream model providers.