gba-planner-7B-v0.2

A planner LLM fine-tuned on synthetic trajectories from an agent simulation. It can be used in ReAct-style LLM agents where planning is separated from function calling. Trajectory generation and planner fine-tuning are described in the bot-with-plan project.

The planner has been fine-tuned on the krasserm/gba-trajectories dataset with a loss over the full sequence (i.e. over prompt and completion). An 8-bit quantized GGUF version of this model is available at krasserm/gba-planner-7B-v0.2-GGUF

Usage example

Load the model and the tokenizer.

import json
import torch
from transformers import (
    AutoModelForCausalLM, 
    AutoTokenizer, 
    BitsAndBytesConfig, 
    GenerationConfig,
)

device = "cuda:0"
repo_id = "krasserm/gba-planner-7B-v0.2"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    quantization_config=bnb_config,
    device_map=device,
)

Define a prompt that contains the user request and past task-observation pairs of the current trajectory (context information).

prompt = """User request:

```
Get the average Rotten Tomatoes scores for DreamWorks' last 5 movies.
```

Context information:

```
Task: Find the last 5 movies released by DreamWorks.
Result: The last five movies released by DreamWorks are "The Bad Guys" (2022), "Boss Baby: Family Business" (2021), "Trolls World Tour" (2020), "Abominable" (2019), and "How to Train Your Dragon: The Hidden World" (2019).

Task: Search the internet for the Rotten Tomatoes score of "The Bad Guys" (2022)
Result: The Rotten Tomatoes score of "The Bad Guys" (2022) is 88%.
```

Plan the next step."""

Then generate a plan for the next step in the trajectory.

instruct_template = "[INST] {prompt} [/INST]"
instruct_prompt = instruct_template.format(prompt=prompt)

input_ids = tokenizer(instruct_prompt, return_tensors="pt", max_length=1024, truncation=True)["input_ids"]
input_ids = input_ids.to("cuda:0")

generation_config = GenerationConfig(
    max_new_tokens=512,
    do_sample=False,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
)

with torch.no_grad():
    result = model.generate(input_ids, generation_config=generation_config)
    result = result[:, input_ids.shape[1] :]

decoded = tokenizer.batch_decode(result, skip_special_tokens=True)
decoded_dict = json.loads(decoded[0])
print(json.dumps(decoded_dict, indent=2))

{
  "context_information_summary": "The last five movies released by DreamWorks are \"The Bad Guys\" (2022), \"Boss Baby: Family Business\" (2021), \"Trolls World Tour\" (2020), \"Abominable\" (2019), and \"How to Train Your Dragon: The Hidden World\" (2019). The Rotten Tomatoes score of \"The Bad Guys\" (2022) is 88%.",
  "thoughts": "Since we have the Rotten Tomatoes score for \"The Bad Guys\", the next logical step is to find the score for the next movie in the list, \"Boss Baby: Family Business\". After obtaining this score, we can proceed to find the scores for the remaining movies in the same manner.",
  "task": "Search the internet for the Rotten Tomatoes score of \"Boss Baby: Family Business\" (2021).",
  "selected_tool": "search_internet"
}

The planner selects a tool and generates a task for the next step. The task is tool-specific and executed by the tool, in this case the search_internet tool, which results in the next observation on the trajectory. If the final_answer tool is selected, a final answer is available or can be generated from the trajectory.

Tools

The planner learned a (static) set of available tools during fine-tuning. These are:

The framework provided by the bot-with-plan project can easily be adjusted to a different set of tools for specialization to other application domains.