---
license: other
---
**NOTE: This "delta model" cannot be used directly.**  
Users have to apply it on top of the original LLaMA weights to get actual LLaMA-30B-toolbench weights.  
Please refer to https://github.com/lm-sys/FastChat#vicuna-weights for instructions.
<br>
<br>

# LLaMA-30B-toolbench

<!-- Provide a quick summary of what the model is/does. -->

LLaMA-30B-toolbench is a 30 billion parameter model used for api based action generation. It is instruction tuned from [LLaMA-30B](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) on api based action generation datasets.

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->

- **Developed by:** [SambaNova Systems](https://sambanova.ai/)
- **Model type:** Language Model
- **Language(s):** English
- **License:** Same as LLaMA model https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md

- **Finetuned from model:** [LLaMA-30B](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md)

### Basic Information

<!-- Provide the basic links for the model. -->
- **Paper**: [Link](https://arxiv.org/abs/2305.16504)
- **Github**: [link](https://github.com/sambanova/toolbench)


## Uses
<details>
<summary>Click to expand</summary>
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

### Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
This model is intended for research use.


### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->


LLaMA-30B-toolbench should NOT be used for purpose other than API based action generation.

</details>


---
## How to Get Started with the Model

<details>
<summary>Click to expand</summary>


### Suggested Inference Parameters
- do_sample: False

### Example Prompts To Try in GPU Tutorial
Prompt 1: 
```
I have the following set of API:\n\n# To set the maximum commute time in minute to your office location, assuming the office location is already defined\nAPI.set_max_commute_time(value: int)\n\n# To set the maximum home size in square feet\nAPI.set_max_square_feet(value: int)\n\n# To set the minimum home price in dollars\nAPI.set_min_price(value: int)\n\n# To set the number of garage(s)\nAPI.set_num_garages(value: int)\n\n# To set home types for search. For home buying, home_types choices are: \"House\", \"Townhouse\", \"Condo\", \"Land\", \"Multi-family\", \"Mobile\", \"Co-op\"; for home renting, home_types choices are: \"House\", \"Townhouse\", \"Condo\", \"Apartment\".\nAPI.select_home_type(home_types: List[str])\n\n# To set the number of balconies\nAPI.set_num_balconies(value: int)\n\n# Submit criterion to get search results. This function should be called after setting all the criterion.\nAPI.search()\n\n# To set the floor number\nAPI.set_floor_number(value: int)\n\n# To set the number of bedroom(s)\nAPI.set_num_beds(value: int)\n\n# To set the number of swimming pool(s)\nAPI.set_num_swimming_pools(value: int)\n\n# To set the maximum home price in dollars\nAPI.set_max_price(value: int)\n\n# To specify whether to search homes for buying or renting. 'value' can be chosen from ['buy', 'rent']. This function must be called after setting the location and before setting any other criteria.\nAPI.set_buy_or_rent(value: str)\n\n# To set the number of bathroom(s)\nAPI.set_num_baths(value: float)\n\n# To set the location for the search area. This function must be called before setting any criteria.\nAPI.set_location(value: string)\n\n# To set the minimum home size in square feet\nAPI.set_min_square_feet(value: int)\n\n-------------\n\nTask: Looking for homes to rent in Santa Clarita with a price range between $110000 and $1753000, a minimum of 1700 square feet, at least 2 balconies, and 3.5 bathrooms.\nAction:\n
```

Prompt 2: 
```
I have the following set of API:\n\n# To set the location for hotel search, given a Loc object. This function must be called if booking type is 'hotels' or 'both'.\nAPI.set_hotel_location(Loc)\n\n# To set the number of hotel rooms to book.\nAPI.set_num_rooms(value)\n\n# To set the location for departure, given a Loc object. This function must be called if booking type is 'trip tickets' or 'both'.\nAPI.set_origin(Loc)\n\n# To select the transportation type from ['flight', 'train', 'bus', 'cruise']. This function must be called if booking type is 'trip tickets' or 'both'.\nAPI.select_transportation(transportation_type)\n\n# To set the return date of the trip, given a Date object. If booking type is 'both' and this function is not called explicitly, 'return_date' will be set to 'hotel_checkout_date' implicitly.\nAPI.set_return_date(Date)\n\n# To set the hotel check-in date, given a Date object. This function must be called if booking type is 'hotels' or 'both'.\nAPI.set_checkin_date(Date)\n\n# To define a date.\ndate = Date(month, day, year)\n\n# To set the departure date of the trip, given a Date object. This function must be called if booking type is 'trip tickets'. If booking type is 'both' and this function is not called explicitly, 'departure_date' will be set to 'hotel_checkin_date' implicitly.\nAPI.set_departure_date(Date)\n\n# To set the location for arrival, given a Loc object. This function must be called if booking type is 'trip tickets' or 'both'.\nAPI.set_destination(Loc)\n\n# To define a location of a given city 'City'.\nlocation = Loc('City')\n\n# To set maximum hotel room price.\nAPI.set_max_room_price(value)\n\n# To set minimum ticket price.\nAPI.set_min_ticket_price(value)\n\n# To select the booking type from ['hotels', 'trip tickets', 'both']. This function must be called before setting any criteria.\nAPI.select_booking_type(booking_type)\n\n# To set minimum hotel room price.\nAPI.set_min_room_price(value)\n\n# To set the number of child tickets to purchase.\nAPI.set_num_children(value)\n\n# To set the number of adult tickets to purchase.\nAPI.set_num_adults(value)\n\n# To select the hotel room type from ['King Bed', 'Queen Bed', 'Double', 'Luxury'].\nAPI.select_room_type(room_type)\n\n# To set maximum ticket price.\nAPI.set_max_ticket_price(value)\n\n# Submit criterion to get search results. This function should be called after setting all the criterion.\nAPI.search()\n\n# To set the hotel check-out date, given a Date object. This function must be called if booking type is 'hotels' or 'both'.\nAPI.set_checkout_date(Date)\n\n-------------\n\nTask: Looking to book 2 adult and 4 child tickets from Stockton to Baltimore by cruise, on 2023-07-29.\nAction:\n
```

</details>

---

## Training Details

<details>
<summary>Click to expand</summary>

### Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The training data is curated for the 8 tasks in ToolBench. See Appendix A of the [paper](https://arxiv.org/abs/2305.16504) for task details and Appendix C.1 for the training data curation details. In total, there are 9704 training samples, organized in all-shot format as described in Appendix C.2. Here is the [download link](https://drive.google.com/file/d/1lUatLGnSVhfy1uVIPEQ7qCoLtnCIXi2O/view?usp=sharing) to the training data.


### Training Procedure 

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

We trained LLaMA-30B-toolbench on 4 80GB A100 gpu's. We started from [LLaMA-30B](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md). We finetuned it on the dataset mentioned above.


### Hyperparameters

- Hardware: A100 GPU
- Optimizer: AdamW
- Grad accumulation: 1
- Epochs: 8
- Global Batch size: 16
- Batch tokens: 16 * 2048 = 32,768 tokens
- Learning Rate: 1e-5
- Learning Rate Scheduler: Fixed LR
- Weight decay: 0.1



</details>



## Acknowledgment

We would like to express our gratitude to the great work done in [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)

## Cite LLaMA-30B-toolbench
```
@misc{xu2023tool,
      title={On the Tool Manipulation Capability of Open-source Large Language Models}, 
      author={Qiantong Xu and Fenglu Hong and Bo Li and Changran Hu and Zhengyu Chen and Jian Zhang},
      year={2023},
      eprint={2305.16504},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```