airtightsquid/TA_GLM4.5_lora_finetune

GLM-4.1V Cryptocurrency Chart Analysis (LoRA Fine-tune)

This model is a LoRA fine-tuned version of zai-org/GLM-4.1V-9B-Thinking for cryptocurrency chart analysis.

Model Details

• —Base Model: GLM-4.1V-9B-Thinking (9B parameter vision-language model)
• —Fine-tuning Method: LoRA (Low-Rank Adaptation)
• —Task: Multi-modal cryptocurrency chart classification (Bullish/Bearish/Neutral)
• —Training Data: 21,521 cryptocurrency chart images with technical indicators

LoRA Configuration

• —Rank (r): 32
• —Alpha: 64
• —Dropout: 0.1
• —Target Modules: qproj, kproj, vproj, oproj
• —Trainable Parameters: 67,787,779 / 10,360,565,251 (0.65%)

Training Results

• —Final Validation Accuracy: 21.81%
• —Final Validation Loss: 0.6934
• —Training Epochs: 5
• —Training Date: 2025-08-18

Vision Configuration

• —Image Size: 336x336
• —Patch Size: 14
• —Grid: [2, 24, 24]

Data

• —25000 normalized TA screenshots from low to med cap solana tokens.

Usage

This model requires custom loading code as it includes both LoRA adapters and a custom classification head.

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoProcessor
import torch

# Load base model and LoRA adapter
config = PeftConfig.from_pretrained("airtightsquid/TA_GLM4.5_lora_finetune")
base_model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16
)
model = PeftModel.from_pretrained(base_model, "airtightsquid/TA_GLM4.5_lora_finetune")

# Load processor
processor = AutoProcessor.from_pretrained(config.base_model_name_or_path, trust_remote_code=True)

# Note: You'll also need to load the custom classifier head (classifier.pt)
# and implement the full GLM41VCryptoModel wrapper for complete functionality

Files Included

• —adapter_model.safetensors: LoRA adapter weights
• —adapter_config.json: LoRA configuration
• —classifier.pt: Custom classification head weights
• —training_config.yaml: Full training configuration

Classes

• —0: Bearish
• —1: Bullish
• —2: Neutral

Limitations

• —Model performance was affected by vision processing issues during training
• —Requires custom wrapper code for full functionality
• —Classification head needs separate loading

License

Please refer to the base model's license: zai-org/GLM-4.1V-9B-Thinking