127 lines
6.8 KiB
Markdown
127 lines
6.8 KiB
Markdown
---
|
|
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:
|
|
- Flexan/Blake-XTM-Arc
|
|
pipeline_tag: text-generation
|
|
---
|
|
# GGUF Files for Blake-XTM-Arc
|
|
|
|
These are the GGUF files for [Flexan/Blake-XTM-Arc](https://huggingface.co/Flexan/Blake-XTM-Arc).
|
|
|
|
| GGUF Link | Quantization | Description |
|
|
| ---- | ----- | ----------- |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q2_K.gguf) | Q2_K | Lowest quality |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.IQ3_XS.gguf) | IQ3_XS | Integer quant |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q3_K_S.gguf) | Q3_K_S | |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.IQ3_S.gguf) | IQ3_S | Integer quant, preferable over Q3_K_S |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.IQ3_M.gguf) | IQ3_M | Integer quant |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q3_K_M.gguf) | Q3_K_M | |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q3_K_L.gguf) | Q3_K_L | |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.IQ4_XS.gguf) | IQ4_XS | Integer quant |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q4_K_S.gguf) | Q4_K_S | Fast with good performance |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q4_K_M.gguf) | Q4_K_M | **Recommended:** Perfect mix of speed and performance |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q5_K_S.gguf) | Q5_K_S | |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q5_K_M.gguf) | Q5_K_M | |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q6_K.gguf) | Q6_K | Very good quality |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.Q8_0.gguf) | Q8_0 | Best quality |
|
|
| [Download](https://huggingface.co/Flexan/Blake-XTM-Arc-GGUF/resolve/main/Blake-XTM-Arc.f16.gguf) | f16 | Full precision, don't bother; use a quant |
|
|
|
|
# Model Card for Blake-XTM Arc
|
|
|
|
Blake-XTM Arc is a 7B large language model used for text generation.
|
|
It was trained to reason and optionally call provided tools.
|
|
|
|
## Model Details
|
|
|
|
### Model Description
|
|
|
|
Blake-XTM Arc is a 7B 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 [CatNyanster-7B](https://huggingface.co/arlineka/CatNyanster-7b) as base model, which was fine-tuned on [Mistral-7B](https://huggingface.co/mistralai/Mistral-7B-v0.1).
|
|
|
|
### Chat Format
|
|
|
|
Blake-XTM Arc 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|>
|
|
```
|
|
|
|
**Warning:** The model seems to bias towards thought process + response, even for short prompts like "Hello," which may cause it to overthink.
|
|
|
|
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': []}}}]
|
|
|
|
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|>
|
|
``` |