AffectGPT Qwen2.5-7B LoRA (epoch 60)

This repository contains only the language-model LoRA tensors extracted from the official AffectGPT epoch-60 checkpoint. It is prepared for a controlled general-capability comparison against the unchanged Qwen/Qwen2.5-7B-Instruct base model.

Provenance

• Upstream code: https://github.com/zeroQiaoba/AffectGPT
• Upstream checkpoint: https://huggingface.co/MERChallenge/AffectGPT
• Source file: checkpoint_000060_loss_0.480.pth
• Paper: AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025)

The upstream checkpoint also contains projectors and multimodal pre-fusion weights. Those tensors are deliberately excluded here. No base-model weights are included.

Adapter configuration

• Base model: Qwen/Qwen2.5-7B-Instruct
• PEFT type: LoRA
• Rank: 16
• Alpha: 32
• Dropout used during training: 0.05
• LoRA parameter count: 40,370,176
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, down_proj, and up_proj in all 28 decoder layers

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-7B-Instruct"
adapter_id = "TianZhenkiseki/AffectGPT-Qwen2.5-7B-LoRA-epoch60"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)

For a fair A/B evaluation, use the same tokenizer, chat template, decoding settings, precision, and benchmark configuration for the base and adapted models. The text-only comparison measures the effect of AffectGPT's LoRA on the language backbone; it does not evaluate the complete multimodal system.

Extraction checks

The converted Safetensors file contains 392 tensors: two LoRA matrices for seven target modules across 28 layers. Layer coverage, tensor count, rank, and parameter count were validated before upload.

Usage notice

The upstream repository includes an Apache-2.0 license file and additionally describes the service as a research preview intended for non-commercial use. Users should review and follow the upstream project's terms and notices.