README.md

---
license: cc-by-sa-4.0
datasets:
- PJMixers-Dev/dolphin-deepseek-1k-think-1k-response-filtered-ShareGPT
- Jofthomas/hermes-function-calling-thinking-V1
language:
- en
base_model:
- microsoft/phi-2
pipeline_tag: text-generation
---
# GGUF Files for Blake-XTM-Arc-3B-V1

These are the GGUF files for [Flexan/Blake-XTM-Arc-3B-V1](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1).

| GGUF Link | Quantization | Description |
| ---- | ----- | ----------- |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q2_K.gguf) | Q2_K | Lowest quality |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.IQ3_XS.gguf) | IQ3_XS | Integer quant |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q3_K_S.gguf) | Q3_K_S |  |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.IQ3_S.gguf) | IQ3_S | Integer quant, preferable over Q3_K_S |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.IQ3_M.gguf) | IQ3_M | Integer quant |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q3_K_M.gguf) | Q3_K_M |  |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q3_K_L.gguf) | Q3_K_L |  |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.IQ4_XS.gguf) | IQ4_XS | Integer quant |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q4_K_S.gguf) | Q4_K_S | Fast with good performance |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q4_K_M.gguf) | Q4_K_M | **Recommended:** Perfect mix of speed and performance |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q5_K_S.gguf) | Q5_K_S |  |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q5_K_M.gguf) | Q5_K_M |  |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q6_K.gguf) | Q6_K | Very good quality |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.Q8_0.gguf) | Q8_0 | Best quality |
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-3B-V1-GGUF/resolve/main/Blake-XTM-Arc-3B-V1.f16.gguf) | f16 | Full precision, don't bother; use a quant |

# Model Card for Blake-XTM Arc 3B (V1)

Blake-XTM Arc 3B (V1) is a 3B large language model used for text generation.
It was trained to reason and optionally call provided tools.

> [!WARNING]  
> **Warning:** this model was finetuned on a **non-instruct** base model. This was an oversight, but it may impact model performance.

## Model Details

### Model Description

Blake-XTM Arc 3B (V1) is a 3B parameter instruct LLM trained to think and optionally call a tool. It only supports using one tool per assistant message (no parallel tool calling).
The model was LoRA fine-tuned with [microsoft/phi-2](https://huggingface.co/microsoft/phi-2) as base model.

### Chat Format

Blake-XTM Arc 3B (V1) uses the ChatML format, e.g.:
```text
<|im_start|>system
System message<|im_end|>
<|im_start|>user
User prompt<|im_end|>
<|im_start|>assistant
Assistant response<|im_end|>
```

### Model Usage

The assistant response can have the following three formats (the contents are examples and were not generated from the model):
1. Only response:
   ```text
   <|im_start|>assistant
   Hello! How may I assist you today?<|im_end|>
   ```
2. Thought process and response:
   ```text
   <|im_start|>assistant
   <|think_start|>The user has greeted me with a simple message. I should think about how to respond to them.

   Since the user sent a simple greeting, I should reply with a greeting that matches their energy.

   Alright, I can reply with a message like 'Hello! How can I help you?'<|think_end|>

   Hello! How may I assist you today?<|im_end|>
   ```
3. Thought process and tool call:
   ```text
   <|im_start|>assistant
   <|think_start|>The user has asked me to find all restaurants near Paris. Hmm... let me think this through thoroughly.

   I can see that I have a tool available called 'find_restaurants', which I might be able to use for this purpose.

   Alright, I think I should use the `find_restaurants` tool to find the restaurants near Paris. For the `city` parameter, I'll use 'Paris', and for the `country` parameter, I'll fill in `France`.

   Okay, I can go ahead and make the tool call now.<|think_end|>

   <|tool_start|>{'name': 'find_restaurants', 'arguments': {'city': 'Paris', 'country': 'France'}}<|tool_end|><|im_end|>
   ```

We recommend using the following system prompts for your situation:
- Only thought process:
  ```text
  You are an advanced reasoning model.
  
  You think between <|think_start|>...<|think_end|> tags. You must think if the user's request involves math or logical thinking/reasoning.
  ```
- Thought process and tool calling:
  ```text
  You are an advanced reasoning model with tool-calling capabilities.

  You think between <|think_start|>...<|think_end|> tags. You must think if the user's request involves math, logical thinking/reasoning, or when you want to consider using a tool.

  # Tools
  You have access to the following tools:
  [{'type': 'function', 'function': {'name': 'convert_currency', 'description': 'Convert currency from one type to another', 'parameters': {'type': 'object', 'properties': {'amount': {'type': 'number', 'description': 'The amount to be converted'}, 'from_currency': {'type': 'string', 'description': 'The currency to convert from'}, 'to_currency': {'type': 'string', 'description': 'The currency to convert to'}}, 'required': ['amount', 'from_currency', 'to_currency']}}}, {'type': 'function', 'function': {'name': 'get_random_joke', 'description': 'Get a random joke', 'parameters': {'type': 'object', 'properties': {}, 'required': []}}}] <\/tools>Use the following pydantic model json schema for each tool call you will make: {'title': 'FunctionCall', 'type': 'object', 'properties': {'arguments': {'title': 'Arguments', 'type': 'object'}, 'name': {'title': 'Name', 'type': 'string'}}, 'required': ['arguments', 'name']}

  To call a tool, write a JSON object with the name and arguments inside <|tool_start|>...<|tool_end|>.
  ```

For responding with a tool response, you can send a message as the `tool` user:
```
<|im_start|>assistant
<|think_start|>The user has asked me to find all restaurants near Paris. Hmm... let me think this through thoroughly.

I can see that I have a tool available called 'find_restaurants', which I might be able to use for this purpose.

Alright, I think I should use the `find_restaurants` tool to find the restaurants near Paris. For the `city` parameter, I'll use 'Paris', and for the `country` parameter, I'll fill in `France`.

Okay, I can go ahead and make the tool call now.<|think_end|>

<|tool_start|>{'name': 'find_restaurants', 'arguments': {'city': 'Paris', 'country': 'France'}}<|tool_end|><|im_end|>
<|im_start|>tool
{'restaurants': [{'name': 'A Restaurant Name', 'rating': 4.5}]}<|im_end|>
```