Model Overview

This repository contains the LoRA adapter for a Qwen3-14B chemical synthesizability classifier, fine-tuned on a Positive/Negative (PN) dataset using QLoRA + focal loss.

Training prompts follow the template:

You are a materials science assistant. Given a chemical composition, answer only with 'P' (synthesizable/positive) or 'N' (non-synthesizable/negative).

Each query was formatted as:

Is the material {composition} likely synthesizable? Answer with P (positive) or N (negative).

Task: Binary classification (P = synthesizable, N = unsynthesizable)
Seed: 3410
Final epoch: 10
Training objective: QLoRA + focal loss (gamma = 2.0, alpha_P = 7.5, alpha_N = 1.0)
Sequence length: 180 tokens
Dataset: train (train_llm_pn.jsonl) / validation (valid_llm_pn.jsonl)
Train/valid datasets are available in data.tar.gz at: https://github.com/evenfarther/Sythesizability_prediction_local_llms

Validation Metrics (Epoch 10; Seed 3410)

Metrics are from logit-based checkpoint evaluation.

Metric	Value
Precision	0.8410
TPR (P Recall)	0.8737
TNR (N Specificity)	0.9773
MCC	0.8372

Dataset Sources

The training/validation splits combine multiple public sources and internal curation:

P/N labelled data from J. Am. Chem. Soc. 2024, 146, 29, 19654-19659 (doi:10.1021/jacs.4c05840).
High-entropy materials data from Data in Brief 2018, 21, 2664-2678 (doi:10.1016/j.dib.2018.11.111).
Additional candidates retrieved via Scopus API queries and manual screening of high entropy materials literature.

VRAM & System Requirements

Base model: unsloth/Qwen3-14B-unsloth-bnb-4bit (Unsloth 4-bit bitsandbytes checkpoint).
Google Colab TPU: this adapter can be used in a Google Colab TPU environment when paired with a TPU-compatible base checkpoint of the same architecture.
Libraries: transformers, peft, bitsandbytes (optionally unsloth).

Limitations & Notes

Evaluation in this project is based on P/N token logits (no generation parsing).
This adapter targets chemical synthesizability classification; generalization outside this domain is not guaranteed.