README.md

---
datasets:
- crownelius/Creative_Writing_ShareGPT_Enhanced
- microsoft/rStar-Coder
- peteromallet/dataclaw-peteromallet
- crownelius/Opus-4.6-Reasoning-3300x
- openbmb/UltraData-Math
- Unreleased/Dataset
base_model:
- trohrbaugh/Qwen3.5-9B-heretic-v2
tags:
- agent
license: apache-2.0
language:
- en
- zh
- ru
- es
- fr
- it
- ja
- ko
- af
- de
- ar
- tr
- is
- pl
- sw
- sv
- nl
- he
- id
- uk
- fa
- pa
- pt
- ms
- fi
- el
---

## This is a Lab Copy for developmental purposes.
If you want to download, please go to [crownelius/Crow-9B-Opus-4.6-Distill-Heretic_Qwen3.5](https://huggingface.co/crownelius/Crow-9B-Opus-4.6-Distill-Heretic_Qwen3.5)


-----------

[<img src="https://huggingface.co/crownelius/Crow-9B-Opus-4.6-Distill-Heretic_Qwen3.5/resolve/main/banner.png" width="350"/>](https://ko-fi.com/abcuo)
CROW HAS REACHED THE FIRST PAGE OF HUGGINGFACE TRENDING MODELS! THANK YOU SO MUCH!!! NUMBER 15 IN THE WORLD!!!


# 🪶 CROW-9B
> **Flagship Intelligence. Featherweight Footprint.**
> *Meticulously distilled from Claude Opus 4.6 into a highly efficient Qwen 3.5 architecture.*

---

**`Architecture:`** `Qwen 3.5` | **`Parameters:`** `9 Billion` | **`Teacher Model:`** `Claude Opus 4.6` | **`Type:`** `Distilled LLM`

---

### 🌟 Model Highlights
* **Distilled Excellence:** Captures the deep reasoning, nuanced formatting, and instruction-following capabilities of Claude Opus 4.6.
* **Highly Agile:** At just 9B parameters, Crow runs efficiently on consumer-grade GPUs and edge devices without sacrificing contextual depth.
* **Qwen 3.5 Backbone:** Inherits robust multilingual support, a massive context window, and structural stability.

--- Generating this model was expensive. You can support this one and future models by tipping. https://ko-fi.com/abcuo

## Available Model files:
- `Qwen3.5-9B-heretic-v2.F16.gguf`
- `Qwen3.5-9B-heretic-v2.Q8_0.gguf`
- `Qwen3.5-9B-heretic-v2.Q5_K_M.gguf`
- `Qwen3.5-9B-heretic-v2.Q4_K_M.gguf`
- `Qwen3.5-9B-heretic-v2.BF16-mmproj.gguf`

## User Guide

### Recommended System Prompt

Default system prompt:

```text
You are Crow, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.

Behavior rules:
- Answer the user's actual request directly.
- Be accurate, complete, and structured.
- Think before answering, but do not get stuck in repetitive loops or meta-commentary.
- If the request is ambiguous or incomplete, state what is missing and make the smallest reasonable assumption needed to continue.
- If the user wants creative writing, preserve tone, continuity, and character consistency.
- If the user wants analysis or technical help, prefer concrete steps, examples, and decisions over fluff.
- Finish with a usable answer, not just planning.
```

Shorter fallback system prompt:

```text
You are Crow. Give direct, useful answers. Keep reasoning concise. Do not loop, do not repeat yourself, and do not pad. If context is missing, say what is missing in one sentence and continue with the best reasonable assumption.
```

### Install & Dependencies

#### LM Studio

1. Install LM Studio from `https://lmstudio.ai/`.
2. Download one of the GGUF files from the Files tab for this repo.
3. Pick a quant based on your hardware:
   - `Q4_K_M`: lowest memory use
   - `Q5_K_M`: best default for most users
   - `Q8_0`: stronger quality, higher memory use
   - `F16`: best quality, highest memory use
4. Load the GGUF in LM Studio.
5. If you need image support in a compatible GGUF runtime, also download the paired `mmproj` file.

Dependencies:

- A current LM Studio build
- Enough disk space for the model and cache
- Enough RAM / VRAM for your selected quant
- Current GPU drivers if you want GPU offload

#### Ollama

1. Install Ollama from `https://docs.ollama.com/quickstart`.
2. Download a Crow 9B GGUF from this repo.
3. Place the GGUF in its own folder.
4. Create a `Modelfile`.
5. Build the model with `ollama create`.

Dependencies:

- A current Ollama release
- Enough disk space for the GGUF and Ollama model store
- Enough RAM / VRAM for your selected quant
- Current GPU drivers if you want GPU acceleration

### Good LM Studio Settings

Recommended starting points:

| Use case | Temperature | Top P | Top K | Repeat penalty | Context | Max tokens |
|---|---:|---:|---:|---:|---:|---:|
| General / reasoning | 0.6 | 0.95 | 20 | 1.05 | 16384 | 4096 |
| Creative writing / roleplay | 0.8 | 0.95 | 40 | 1.02 | 16384-32768 | 4096-8192 |

Notes:

- Start with `Q5_K_M` unless you have a reason to use a different quant.
- Lower temperature if you see rambling or unstable reasoning.
- Do not max out context by default. Larger contexts cost more memory and can make long chats less stable on weaker hardware.

### Good Ollama Settings

Example `Modelfile`:

```text
FROM ./Qwen3.5-9B-heretic-v2.Q5_K_M.gguf

PARAMETER num_ctx 16384
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER repeat_penalty 1.05
PARAMETER repeat_last_n 256

SYSTEM """
You are Crow, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.
Answer directly, stay coherent, avoid repetitive thinking loops, and finish with a complete answer.
If context is missing, identify the gap briefly and continue with the best reasonable assumption.
"""
```

Build and run:

```bash
ollama create crow-9b -f Modelfile
ollama run crow-9b
```

For more creative outputs, raise `temperature` to `0.8`, raise `top_k` to `40`, and reduce `repeat_penalty` slightly to `1.02`.

### If You Hit a Thinking Loop

If the model starts looping inside `<think>` tags, repeating analysis, or stalling:

1. Stop generation.
2. Retry with a lower temperature, ideally `0.4` to `0.6`.
3. Increase repeat penalty slightly, for example from `1.05` to `1.08`.
4. Add an instruction like:

```text
Answer directly. Keep reasoning brief. Do not repeat analysis. Give the final answer.
```

5. Start a fresh chat if the current conversation has become unstable.
6. If the problem happens mostly on lower quants, move up to `Q8_0` or `F16`.

### If the Prompt Is Incomplete or the Output Cuts Off

If the prompt is incomplete or malformed:

1. Clean it before sending it.
2. Remove broken tags, clipped instructions, or half-finished bullets.
3. If only partial context is available, prepend:

```text
If context is missing, state your assumptions briefly and continue with the most likely intended task.
```

If the output cuts off:

1. Increase max tokens.
2. Ask:

```text
Continue from the last complete sentence. Do not restart or summarize. Continue exactly where you stopped.
```

3. If it still restarts instead of continuing, begin a fresh chat and resend the prompt with the desired output format stated more explicitly.

### Prompting Tips

- State the exact deliverable: list, table, code, rewrite, draft, critique, or decision.
- For coding, specify language, runtime, and expected input/output.
- For creative writing, specify tone, genre, constraints, and point of view up front.
- For high-control tasks, say whether you want concise output, full reasoning, or final answer only.

This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)