Modelpublic

xoykor/Ornith-1.5-9B-DSH-Agentic-GPT-5.6-Sol-Distill

sourceHugging Facemitupdated 6d agoView on Hugging Face

3likes506downloads

Model Card

Ornith-1.5-9B-DSH-Agentic-GPT-5.6-Sol-Distill

Ornith-1.5-9B-DSH-Agentic-GPT-5.6-Sol-Distill is an unofficial community fine-tune of ornith-ai/Ornith-1.5-9B, specialized for agentic workflows, tool use, coding, and DSH-style harnesses.

The specialization data was produced with GPT-5.6 Sol as a teacher and was designed to encourage a more action-oriented agent loop:

Understand → Act → Observe → Correct → Finish

The goal is not to replace Ornith's general capabilities, but to make its behavior more useful in tool-heavy environments where the model is expected to act instead of spending excessive tokens planning.

Highlights

—Base model: ornith-ai/Ornith-1.5-9B
—Fine-tuning method: LoRA
—Teacher: GPT-5.6 Sol
—Primary focus: agentic tool use and coding
—GGUF quantization: Q5_K_M
—Vision projector: F16 mmproj
—Multimodal capability inherited from the base Ornith model
—Designed for local inference and OpenAI-compatible agent harnesses
—Intended for DSH / DeepSeek Harness-style workflows

What was trained

The specialization dataset focuses on behaviors such as:

—using tools early when a tool is actually required;
—choosing an appropriate tool instead of overthinking;
—interpreting tool outputs before taking the next action;
—recovering from failed tool calls by changing strategy;
—avoiding repeated calls that already failed;
—avoiding unnecessary tool calls;
—handling multi-step tasks;
—writing and modifying code;
—using Python and common libraries when appropriate;
—working with files, repositories, shells, and development workflows;
—asking for clarification only when required information is genuinely missing;
—keeping operational reasoning concise;
—stopping when the task has already been completed.

The intended behavior is a practical agent loop rather than long speculative planning.

Model lineage

text

ornith-ai/Ornith-1.5-9B
          │
          │ LoRA fine-tuning
          │ teacher: GPT-5.6 Sol
          ▼
Ornith-1.5-9B-DSH-Agentic
          │
          │ merge + GGUF conversion
          ▼
Q5_K_M language model
          +
F16 multimodal projector

Files

The GGUF release uses two components:

text

Ornith-1.5-9B.Q5_K_M.gguf
Ornith-1.5-9B.F16-mmproj.gguf

The first file contains the quantized language model.

The mmproj file contains the multimodal projector required for image input in compatible GGUF runtimes.

For vision, keep the language-model GGUF and the matching mmproj available to the same runtime.

LM Studio

Import the Q5_K_M GGUF into LM Studio and keep the corresponding F16 mmproj in the same model directory.

The language model is the main model to load. The mmproj is an auxiliary vision component, not a second chatbot model.

If LM Studio does not automatically associate the projector, verify that the model and projector are in the same model directory and that the selected runtime supports the model's multimodal architecture.

llama.cpp

With a sufficiently recent llama.cpp build that supports the architecture, the model can be served with the language model and multimodal projector together.

Example:

bash

llama-server \
  -m Ornith-1.5-9B.Q5_K_M.gguf \
  --mmproj Ornith-1.5-9B.F16-mmproj.gguf \
  -c 32768 \
  --port 8000

For agentic usage, expose the model through an OpenAI-compatible endpoint and use a harness that preserves the model's chat template and tool-call formatting.

DSH / agentic usage

This fine-tune is intended primarily for environments where the model can interact with real tools:

text

User request
    ↓
Model
    ↓
Tool call
    ↓
Tool result
    ↓
Model evaluates result
    ↓
Next action or final answer

It is specifically tuned to prefer useful actions over unnecessary internal planning.

Agentic behavior depends heavily on the runtime, chat template, tool schema, parser, context size, and harness configuration.

Vision

Ornith-1.5-9B is multimodal, and this release keeps the GGUF multimodal projector as a separate F16 mmproj file.

The agentic fine-tuning was primarily targeted at language/tool behavior. The vision stack is inherited from the base model rather than being the main target of this specialization.

Quantization

The primary release uses Q5_K_M.

This was selected as a compromise between:

—model size;
—local inference speed;
—VRAM usage;
—preservation of the fine-tuned behavior.

Quantization can still cause small behavioral differences compared with the merged higher-precision model.

Benchmarks

Benchmarking of this fine-tune is still in progress.

The intended comparison is:

text

Ornith-1.5-9B base
        vs.
Ornith-1.5-9B-DSH-Agentic-GPT-5.6-Sol-Distill

under the same runtime, quantization, context length, harness, and sampling settings.

Agentic benchmarks such as Terminal-Bench, Toolathlon, and MCP-style evaluations are particularly relevant to the purpose of this model.

Until reproducible results are published here, no claim is made that this fine-tune matches or exceeds the official Ornith benchmark scores.

Intended use

This model is intended for:

—local AI agents;
—coding assistants;
—terminal and shell workflows;
—repository work;
—file manipulation;
—tool-calling experiments;
—DSH-style agent harnesses;
—multi-step automation;
—research into compact agentic models;
—multimodal local inference where supported.

Limitations

This model inherits limitations from the base model and from the teacher-generated specialization data.

In particular:

—tool calls can still be incorrect or unnecessary;
—the model can hallucinate tool outputs if the runtime does not actually execute tools;
—agentic performance depends strongly on the harness;
—Q5KM quantization can differ from higher-precision inference;
—multimodal behavior depends on correct loading of the matching mmproj;
—the model can produce incorrect code or factual claims;
—successful execution should be verified when the model performs consequential actions.

About the distillation

“GPT-5.6 Sol Distill” in the repository name describes the teacher used to generate specialization data.

This is not a GPT-5.6 model, does not contain GPT-5.6 weights, and is not an official OpenAI model.

It remains an Ornith-1.5-9B derivative fine-tuned on teacher-generated examples.

Attribution

Base model:

ornith-ai/Ornith-1.5-9B

Fine-tune / release:

xoykor/Ornith-1.5-9B-DSH-Agentic-GPT-5.6-Sol-Distill

Thanks to the Ornith team for releasing the base model and to the open-source tooling ecosystem used for training and GGUF inference.

License

This derivative follows the MIT License used by the base Ornith-1.5-9B model.

See the repository license file for the full license text.

Status

This is an experimental agentic fine-tune.

Feedback, benchmark results, runtime compatibility reports, and reproducible agent evaluations are welcome.