CoolFace
Modelpublic

archon159/LoGo-loras-collection

sourceHugging Facemitupdated 5mo agoView on Hugging Face
1likes
Model Card

LoGo LoRA Collection

A collection of LoRA adapters used in [LoGo: LoRA on the Go](https://github.com/archon159/LoGo) (ACL 2026).

LoGo dynamically selects and merges the most relevant LoRA adapters at inference time for each input instance. This collection provides the 260 task-specific adapters per base model that LoGo selects from.

Contents

Each base model has its own subfolder containing 260 LoRA adapters trained on FlanV2 tasks:

LoGo-loras-collection/
├── Llama-3.1-8B-loras-flanv2/
│   ├── flanv2.<task_name>/
│   │   ├── adapter_config.json
│   │   └── adapter_model.safetensors
│   └── ...  (260 adapters)
├── Qwen2.5-7B-loras-flanv2/       (260 adapters)
└── deepseek-llm-7b-base-loras-flanv2/  (260 adapters)

The full list of 260 FlanV2 tasks is provided in `flanv2_task_list`.

Training Details

All adapters share the same LoRA configuration:

HyperparameterValue
Rank (r)16
Alpha (lora_alpha)16
Dropout0.05
Target modulesq_proj, v_proj
Task typeCAUSAL_LM

Usage

With LoGo

Follow the instructions in the LoGo repository. Adapters are downloaded automatically.

bash
python3 main.py --base_model Llama-3.1-8B --dataset bbh.boolean_expressions --gpu 0

Standalone (via PEFT)

Individual adapters can be loaded directly using the PEFT library:

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

model = PeftModel.from_pretrained(
    base_model,
    "archon159/LoGo-loras-collection",
    subfolder="Llama-3.1-8B-loras-flanv2/flanv2.ai2_arc_ARC-Challenge",
)