pip-code-bandit

Objective

Given a goal and tools, can AI intelligently use the tools to reach the goal?
What if it has a meagre 1.3b params/neurons akin to that of an owl? Can it follow instructions and plan to reach a goal?
It can!
Releasing pip-code-bandit and pipflow
A model and a library to manage and run goal-oriented agentic system.

Model attributes

-- number of params ~ 1.3b [2.9 Gb GPU memory footprint]
-- sequence length ~ 16.3k [Can go higher but will show performance degradation]
-- license - apache 2.0
-- instruction following , RL tuned.
-- tasks:
1. complex planning(plan) of sequential function calls | a list of callables and goal
2. corrected plan | feedback instructions with error
3. function calling | doc or code and goal
4. code generation | plan and goal
5. code generation | goal
6. doc generation | code
7. code generation | doc
8. file parsed to json | any raw data
9. sql generation | schema, question, instructions and examples

How did we build it?

We used a simulator to simulate environments where the model could play games to achieve goals, given a set of actions available to it. All the model could do was find the right action and config to incur a positive reward. The reward policy is around the concept of a model going to a stable state of zero net sum reward for both good and bad behaviour. In this setup, the model, which was pre-trained on code, function documentation, and similar OS datasets, was RL-tuned for reliability and instruction-following.

License

complete open-sourced - apache 2.0. License

Usage

NOTE:

If you wish to try this model without utilizing your GPU, we have hosted the model on our end. To execute the library using the hosted model, initialize the generator as shown below:

pip3 install git+https://github.com/PipableAI/pipflow.git

from pipflow import PipFlow

generator = PipFlow()

We have hosted the model at https://playground.pipable.ai/infer. Hence, one can also make a POST request to this endpoint with the following payload:

{
    "model_name": "PipableAI/pip-code-bandit",
    "prompt": "prompt",
    "max_new_tokens": "400"
}

curl -X 'POST' \
  'https://playground.pipable.ai/infer' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/x-www-form-urlencoded' \
  -d 'model_name=PipableAI%2Fpip-code-bandit&prompt="YOUR PROMPT"&max_new_tokens=400'

Alternatively, you can directly access the UI endpoint at https://playground.pipable.ai/docs#/default/infer_infer_post.

Library Usage

To directly use the model's capabilities without putting extra effort into schemas and prompts, try to use pipflow.

For detailed usage, refer to the colab_notebook

Model Usage

pip install transformers accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from accelerate import Accelerator
model =AutoModelForCausalLM.from_pretrained("PipableAI/pip-code-bandit",torch_dtype=torch.bfloat16,device_map="auto")
tokenizer = tokenizer = AutoTokenizer.from_pretrained("PipableAI/pip-code-bandit")
new_tokens = 600
prompt = """
<question>
Generate a python function for adding two numbers.
</question>
<code>
"""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=new_tokens)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
response = response.split("<code>")[1].split("</code>")[0]

print(response)

Prompt

prompt = f"""<example_response>{--question , --query}</example_response><function_code>{code}</function_code>
<question>Give one line description of the python code above in natural language.</question>
<doc>"""

prompt = f"""<example_response>{example of some  --question: , --query}</example_response><schema>{schema with cols described}</schema>
<question>Write a sql query to ....</question>
<sql>"""

Team

Avi Kothari, Gyan Ranjan, Pratham Gupta, Ritvik Aryan Kalra, Soham Acharya