GPT-oss-MOFinder-Leave1FieldOut
This repository shows the ablation on removing 1-out-of-8 field to train GPT opensource base model. We reports the performance of LLMs after domain-specific LoRA finetuning by different settings of chemical information missing.
Some syntheses are more sensitive on one single variable than others.
This drop-single-field branch provides the answer whether the model can maintain strong performance without relying on single-field P→N flips during training.
In this ablation training set, we remove such failure experiment records occurs by changing only a single field from the experimental conditions of a successful synthesis.
Result
| Train set drop | Inference | Test size | Accuracy | Precision |
|---|---|---|---|---|
| drop 1-field 'flipped' N entrees | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 14/22 | 0.5789 | |
| matched random drop N | ||||
| seed 101 | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 16/22 | 0.6471 | |
| seed 202 | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 15/22 | 0.6429 | |
| seed 303 | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 16/22 | 0.6667 | |
| seed 404 | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 14/22 | 0.5789 | |
| seed 505 | Holdout | 1320P+ 1275N | 0.73141 | 0.77824 |
| 22 questions | 11P + 11N | 16/22 | 0.6471 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "GPT-oss-MOF-FieldLeave1Out"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
To go throught the quick start details for reprduction, see the official released MOFinder main checkpoint which is well-trained on full-field PN data.
Citation and license
When using these weights, cite MOFinder, identify StarLiu714/GPT-oss-MOF and the Hub revision used, and reference the GPT-oss-20B base model.
This release is licensed under the GNU General Public License v3.0; see LICENSE.
The underlying GPT-oss-20B model retains its upstream Apache-2.0 license, reproduced in LICENSE.base-model. This release includes MOF fine-tuning changes merged into the base model and a BF16 weight export.
MOFinder codebase is separately licensed under MIT.