This repository contains the LoRA adapter for the GPT-OSS 20B chemical synthesis classifier fine-tuned with focal loss. Training prompts follow the template:

As chief synthesis scientist, judge {compound} critically. Avoid bias. P=synthesizable, U=not:

The base checkpoint remains in Unsloth's MXFP4 4-bit format (unsloth/gpt-oss-20b-unsloth-bnb-4bit). Attaching this adapter restores the best validation performance obtained during training.

Task: Binary classification (P = synthesizable, U = unsynthesizable)
Training Objective: QLoRA with focal loss (gamma = 2.0, alpha_P = 8.12, alpha_U = 1.0)
Sequence Length: 180 tokens
Dataset: 316,442 train (train_pu_hem.jsonl) / 79,114 validation (validate_pu_hem.jsonl) samples (11% P / 89% U)
Hardware: NVIDIA RTX A6000 48GB
Adapter Size: 127 MB

Validation Metrics (Epoch 4 - Best)

Metric	Value
TPR (P Recall)	0.9105
TNR (U Specificity)	0.9249

Dataset Sources

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

P/U 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.

After de-duplication across all sources, approximately 2,560 unique compositions were appended to the base dataset. The combined corpus contains 316,442 training samples and 79,114 validation samples with 11 percent P and 89 percent U class balance.

VRAM & System Requirements

GPU VRAM: 12-14 GB (T4 / L4 / A10G class works). The base model loads in 4-bit MXFP4 via Unsloth.
RAM: >=16 GB recommended for tokenization and batching.
Libraries: unsloth>=2025.9.6, transformers>=4.56.0, peft>=0.17.1, bitsandbytes>=0.47.0.
CPU-only inference is not supported with this adapter (MXFP4 requires GPU).

Limitations & Notes

MXFP4 4-bit weights require a GPU; CPU-only inference is not supported.
The dataset targets chemical synthesizability. Generalization outside this domain is not guaranteed.

evenfarther/gpt-20b-chemical-synthesis-adapter-high-TPR

GPT-OSS 20B Chemical Synthesis Classifier - LoRA Adapter

Model Overview

Validation Metrics (Epoch 4 - Best)

Dataset Sources

VRAM & System Requirements

Limitations & Notes

evenfarther/
gpt-20b-chemical-synthesis-adapter-high-TPR