File size: 2,441 Bytes
39fe4e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63e6d5b
39fe4e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2eb89d4
39fe4e3
 
 
 
63e6d5b
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---

license: apache-2.0
base_model: wangzhang/Qwen3.5-4B-abliterated
pipeline_tag: text-generation
library_name: gguf
tags:
- gguf
- qwen3.5
- code
- coding
- abliterated
- ollama
language:
- en
---


# YokoCoder-4B

YokoCoder-4B is a compact, locally runnable coding assistant built on the abliterated Qwen3.5 4B model. It was fine-tuned for coding, further trained on Flappy Bird examples, and continued for five more epochs on a new coding dataset. It builds on the author's checkpoint that passed the Flappy Bird game-generation test. This five-epoch release was tested in Ollama and passed a three-case Python function smoke test, offering practical coding value for its size.

## Model details

- **Format:** GGUF, Q4_K_M
- **Parameters:** 4B
- **Download size:** about 2.7 GB
- **Base:** [wangzhang/Qwen3.5-4B-abliterated](https://huggingface.co/wangzhang/Qwen3.5-4B-abliterated), derived from Qwen3.5-4B
- **Training:** coding fine-tune, three additional epochs on Flappy Bird examples, then five additional epochs on 12 coding examples (28 training windows)
- **License:** Apache 2.0

## Run with Ollama

Download the files in this repository, then run the following from their directory:

```bash

ollama create yokocoder-4b -f Modelfile

ollama run yokocoder-4b

```

The included Modelfile configures a 16,384-token context and sets `num_predict -1` for generation without a fixed output-token cap. The GGUF can also be used with other compatible Qwen3.5 GGUF runtimes.

## Attribution

YokoCoder-4B is a fine-tuned and quantized derivative of [Qwen3.5-4B-abliterated](https://huggingface.co/wangzhang/Qwen3.5-4B-abliterated) by Wangzhang Wu, based on [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B). The original model card and notices are included in [UPSTREAM_MODEL_CARD.md](UPSTREAM_MODEL_CARD.md), and the license is in [LICENSE](LICENSE).

## Latest training update

The October 29, 2026 update continues the existing YokoCoder checkpoint for five epochs, completing 20 optimizer steps with training loss 1.0559. The Q4_K_M file contains the merged updated weights. The continued-training LoRA adapter and its training records are included under `adapters/dataset4-5epochs/`. Export hashes and provenance are in `export_manifest_5epochs.json`.

The local generation check and model identity are recorded in [ollama_smoke_test_5epochs.json](ollama_smoke_test_5epochs.json).