ProjAtlas-xLAM2-8B-FunctionCalling-lora

LoRA adapter fine-tuned from Salesforce/Llama-xLAM-2-8b-fc-r for the ProjAtlas platform's neuron selection and brain visualization agents.

ProjAtlas is an AI-assisted multi-agent platform for interactive exploration of single-neuron projectomes across mouse and macaque brains. This adapter powers the function-calling component that translates natural-language instructions into structured tool calls (e.g. filtering neurons by soma location or projection target, adjusting the interactive 3D viewport).

Training code, dataset, and evaluation scripts are in the function-calling/ directory of the ProjAtlas repository.

⚠️ Scope and Limitations

This is not a general-purpose function-calling model. It is fine-tuned specifically on the ProjAtlas platform's own tool schema (three toolsets: Mouse, Macaque, and the shared Neuroviz visualization toolset). It is intended to reproduce this paper's reported results and is unlikely to perform well on function-calling tasks or tool schemas outside this dataset. See the dataset and data/train.json for the exact tool definitions this adapter was trained on.

Results

Held-out test set, 865 samples:

Compared with 34.0% for the unmodified base model and 81.0% for the strongest of five general-purpose baselines evaluated (Gemini 3 Pro Preview) on the same test set.

Training Details

• Base model: Salesforce/Llama-xLAM-2-8b-fc-r
• Method: LoRA supervised fine-tuning (rank=8, alpha=16, target=all linear layers), via LLaMA-Factory
• Training data: 3,504 samples, curated conversational queries paired with structured function-calling tool schemas (see repository for full dataset)
• Epochs: 3, learning rate 1e-4, cosine schedule, effective batch size 8, bf16

Usage

Requires peft and transformers. Load the base model and apply this adapter:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("Salesforce/Llama-xLAM-2-8b-fc-r")
tokenizer = AutoTokenizer.from_pretrained("Salesforce/Llama-xLAM-2-8b-fc-r")
model = PeftModel.from_pretrained(base_model, "zhang-manyi/ProjAtlas-xLAM2-8B-FunctionCalling-lora")

For merging the adapter into the base model and deploying with vLLM, see the "Merge LoRA Weights" and "Local Deployment with vLLM" sections of the repository README.

License

This adapter is derived from Salesforce/Llama-xLAM-2-8b-fc-r and is licensed under CC-BY-NC-4.0 (non-commercial, research use only), subject also to the Meta Llama Community License. This differs from the Apache-2.0 license covering the training/evaluation code in the ProjAtlas repository.

Citation

See the ProjAtlas GitHub repository for the current citation.