metadata
library_name: transformers
license: apache-2.0
datasets:
- ncauchi1/general_cv_dataset_1
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
Model Card for Model ID
Second version of VLLM fine tuned to answer general questions about cyclic voltammographs. Evaluated on bxw315-umd/general-cv-questions
Uses
Used to answer general multiple choice questions about cyclic voltammogram graphs
Training Details
Trained on ncauchi1/general_questions_dataset with 12k samples. Logs found here: [https://wandb.ai/ncauchi-university-of-maryland/huggingface/runs/i3455num/logs]
Training dataset consists of a mix of data:
10k Pointing samples
- Given graphs with a range 2-4 CVs and points out all peaks in voltage/currents
- graphs do not have legends
2k Question samples
- Given question, graph, and four options
- gives reasononing and correct answer
- graphs are catagorized into templates and questions/reasoning are generated based on graph template
- Graphs are generated from raw data gathered by me, consisting of CV's of Ferrocene and Tryptophan in PBS with concentrations of 0uM, 100uM and 200uM.
Evaluation
35% ± 5 chance to answer correct Evaluation done on bxw315-umd/general-cv-questions, with an 15% increase in performance over base model (35% chance to answer correct)
Error analysis reveals trying to identify features and connecting them to reasoning. Error analysis shows model still has trouble percieving graphs correctly