Dizel 1.5

Experimental · ~455M parameters · F16 GGUF · 2,048-token context

Dizel 1.5 is a small, from-scratch causal language model by D4niel-dev, packaged for local inference with Ollama and llama.cpp. It is a research preview, not a production assistant.

Architecture note: the GGUF uses the qwen2 architecture identifier so it can run in Ollama and llama.cpp. Dizel 1.5 is not a Qwen model or a Qwen fine-tune. Qwen2.5-0.5B-Instruct was used only as a teacher for LoRA distillation; the student weights, vocabulary, and identity are Dizel’s.

Quick start

ollama run D4niel-dev/Dizel-1.5

The first run downloads the F16 model (~1.1 GB). This release was tested with Ollama 0.34.0; use a current release, since older versions may not support the multi-turn .Messages template shown below.

>>> Hello!
>>> Explain what a for loop does in one paragraph.

If the model starts repeating, press Ctrl+C, lower temperature, and raise repeat_penalty (see Settings).

At a glance

Ollama displays 557.18M parameters because the tied input/output embedding is serialized twice in the GGUF. The model has 454,779,904 unique parameters (~455M). The source GGUF files are 1.04 GiB (1,116,875,808 bytes) each; the published Ollama package is about 1.1 GB.

Intended use

Dizel 1.5 is intended for:

• Local experiments with a small from-scratch language model.
• Testing the Dizel training, conversion, and Ollama packaging pipeline.
• Offline demos, education, and prompt-format experiments.
• Comparing checkpoints and generation settings.

It is not intended for:

• Production or customer-facing use.
• Medical, legal, financial, safety-critical, or other high-stakes decisions.
• Autonomous agents, tool execution, or unattended code generation.
• Processing personal, confidential, or regulated data.
• Any use where a wrong or fabricated answer could cause harm.

Prompt format

Dizel 1.5 was trained with literal role markers. The format is:

<|system|>You are Dizel, a helpful assistant.<|end|><|user|>Hello!<|end|><|assistant|>

The assistant response should end with <|end|>. The special tokens are:

Do not use ChatML (<|im_start|>), Llama headers, or [INST] formatting. Those formats were not used during training and will produce poor results.

Build locally from a GGUF

If you have an F16 GGUF, create an Ollama model with the template and parameters below. The FROM path is relative to the Modelfile.

FROM ./Dizel-1.5-f16.gguf

TEMPLATE """{{- if .System }}<|system|>{{ .System }}<|end|>{{ end }}
{{- range .Messages }}{{ if eq .Role "user" }}<|user|>{{ .Content }}<|end|>{{ else if eq .Role "assistant" }}<|assistant|>{{ .Content }}<|end|>{{ end }}{{ end }}
{{- if .Prompt }}<|user|>{{ .Prompt }}<|end|>{{ end }}<|assistant|>"""

PARAMETER stop <|end|>
PARAMETER stop </s>
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 2048

Save it as Modelfile, then run:

ollama create dizel-1.5 -f Modelfile
ollama run dizel-1.5

The published Ollama tag uses this exact template and these parameters. If you build from one of the repository GGUFs, replace ./Dizel-1.5-f16.gguf with the actual filename (Dizel-SFT.gguf or Dizel-Distill.gguf).

Settings

These are the published defaults plus practical tuning tips:

The model has a known repetition failure mode. A repetition penalty can help, but setting it too high may make the answer worse. Start conservatively and compare outputs at a fixed seed.

API examples

Python, using the Ollama Python library:

from ollama import chat

response = chat(
    model="D4niel-dev/Dizel-1.5",
    messages=[
        {"role": "system", "content": "You are Dizel. Answer concisely."},
        {"role": "user", "content": "Hello!"},
    ],
)
print(response.message.content)

cURL:

curl http://localhost:11434/api/chat -d '{
  "model": "D4niel-dev/Dizel-1.5",
  "messages": [{"role": "user", "content": "Hello!"}],
  "stream": false
}'

For raw completions, send the prompt format above to /api/generate.

Files and release status

This release is F16 only. Q4_K_M, Q8_0, and other quantized tags are not published. Do not assume a Q4/Q8 artifact exists; third-party quantizations are unofficial and have not been verified for this release.

A technical note for people inspecting GGUF metadata: some older local F16 builds contain a stale general.file_type=2 value. The tensor data is F16/F32, and the published Ollama tag reports F16. That metadata value does not indicate a Q4 release.

Training

Dizel 1.5 is a dense model trained from scratch on a curated corpus of roughly 378M tokens (code, math, general text, and safety data).

• Tokenizer: 100K SentencePiece BPE with byte fallback.
• Context: 2,048 tokens; the model was not trained for longer contexts.
• Alignment: SFT and LoRA distillation. DPO remains experimental and is not part of this release.
• Training data: D4niel-dev/Dizel-Datasets.
• Teacher: Qwen2.5-0.5B-Instruct (Apache-2.0), used for behavior/quality distillation only. Dizel keeps its own identity and does not share Qwen’s weights.

Evaluation and known limitations

No standardized benchmark results are published for Dizel 1.5. It is a research preview, and the internal evaluation is qualitative.

Internal prompt checks found that simple greetings can work, but arithmetic, grammar correction, list following, long-form answers, and repetition are unreliable. The model can produce off-topic, incomplete, repetitive, or fabricated text. It is not comparable to Qwen2.5-0.5B-Instruct or other instruction-tuned models.

Known limitations:

• 2,048-token context; long documents are truncated.
• English- and code-oriented training data; limited multilingual ability.
• No vision, tool calling, function calling, or reliable structured output.
• Weak arithmetic and grammar; frequent repetition on open-ended prompts.
• No safety alignment guarantee. It may produce harmful or biased content.
• Only F16 is published; there are no official Q4/Q8 tags.

Treat every output as a draft. Verify facts, code, and instructions before using them. If you need reliable answers, use a larger instruction-tuned model.

License

• Model weights: Apache-2.0.
• Dizel repository code: MIT.
• Training data: source datasets retain their original licenses; see the dataset card.
• Teacher: Qwen2.5-0.5B-Instruct is Apache-2.0. Dizel is an independent project and is not affiliated with or endorsed by Alibaba or the Qwen team.
• Dizel name and banner: D4niel-dev.

If you redistribute the model, keep the license and attribution notices.

Links

• Ollama model page
• Hugging Face model card
• Training dataset
• GitHub repository
• Issues and feedback

Citation

@misc{dizel15,
  author       = {D4niel-dev},
  title        = {Dizel 1.5},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/D4niel-dev/Dizel-1.5}}
}

Acknowledgements

Thanks to the Qwen team for the teacher model, SentencePiece for the tokenizer tooling, and the llama.cpp and Ollama projects for local inference.

Model card updated 2026-10-21. Dizel 1.5 is an experimental release; behavior and packaging may change without notice.