Base model: Qwen/Qwen2.5-1.5B-Instruct
Method: QLoRA (r=16, alpha=32, dropout=0.05)
Task: Causal LM — intent routing and structured action planning for the EVd3x platform
License: Apache 2.0

Training Data

3,816 training examples across 15 intent categories
424 eval examples
Sources: EVd3x instruction corpus (Phase 1) covering miRNA targets, pathway enrichment, cell specificity, cell communication, disease associations, EV evidence, and ligand-receptor analysis

Intent categories covered

network_lookup · ev_evidence · pathway · disease · ppi · cell_specificity · cell_communication · lr_analysis · reverse_disease · reverse_cell_comm · figure_export · explain_general · explain_term · collection_lookup · localization

Training Procedure

Epochs: 3
Batch size: 4 (effective batch 16 with grad accum 4)
Learning rate: 2e-4 with 3% warmup
Max seq length: 2048
LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Hardware: Google Colab T4 (free tier)

How to Use

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "YOUR_HF_USERNAME/evd3x-agent-lora-qwen15b")

inputs = tokenizer("What pathways involve SNCA in Parkinson's disease?", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Intended Use

This adapter is designed for use within the EVd3x platform as the LLM backbone of the research assistant agent. It routes natural language queries to structured action plans that drive the EVd3x web UI.

Citation

If you use this model, please cite the EVd3x platform.

jonbrees/evd3x-agent-lora-qwen15b

EVd3x-Agent LoRA — Qwen2.5-1.5B-Instruct

Model Details