jiazhengli
commited on
Commit
•
8f97e33
1
Parent(s):
0ffb2d3
Update README.md
Browse files
README.md
CHANGED
@@ -5,43 +5,61 @@ tags:
|
|
5 |
- llama-factory
|
6 |
- lora
|
7 |
- generated_from_trainer
|
8 |
-
base_model:
|
|
|
|
|
9 |
model-index:
|
10 |
- name: sft_trained_woaqa_mixtral_dpo
|
11 |
results: []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
---
|
13 |
|
14 |
-
|
15 |
-
should probably proofread and complete it, then remove this comment. -->
|
16 |
|
17 |
-
|
18 |
|
19 |
-
|
20 |
-
|
21 |
-
- Loss: 2.6468
|
22 |
-
- Rewards/chosen: 14.3798
|
23 |
-
- Rewards/rejected: 12.6853
|
24 |
-
- Rewards/accuracies: 0.6426
|
25 |
-
- Rewards/margins: 1.6945
|
26 |
-
- Logps/rejected: -209.4146
|
27 |
-
- Logps/chosen: -167.4447
|
28 |
-
- Logits/rejected: -0.8532
|
29 |
-
- Logits/chosen: -0.8500
|
30 |
-
|
31 |
-
## Model description
|
32 |
-
|
33 |
-
More information needed
|
34 |
|
35 |
## Intended uses & limitations
|
36 |
|
37 |
-
|
38 |
|
39 |
## Training and evaluation data
|
40 |
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
## Training procedure
|
44 |
|
|
|
|
|
45 |
### Training hyperparameters
|
46 |
|
47 |
The following hyperparameters were used during training:
|
|
|
5 |
- llama-factory
|
6 |
- lora
|
7 |
- generated_from_trainer
|
8 |
+
base_model:
|
9 |
+
- jiazhengli/Mixtral-8x7B-Instruct-v0.1-QLoRA-Assessment-Rationale-sft
|
10 |
+
- mistralai/Mixtral-8x7B-Instruct-v0.1
|
11 |
model-index:
|
12 |
- name: sft_trained_woaqa_mixtral_dpo
|
13 |
results: []
|
14 |
+
datasets:
|
15 |
+
- jiazhengli/Rationale_MCTS
|
16 |
+
- jiazhengli/Synthetic_Rationale
|
17 |
+
language:
|
18 |
+
- en
|
19 |
+
metrics:
|
20 |
+
- accuracy
|
21 |
+
- f1
|
22 |
---
|
23 |
|
24 |
+
# Mixtral-8x7B-Instruct-v0.1-QLoRA-Assessment-Rationale-dpo
|
|
|
25 |
|
26 |
+
The model trained with w/o private data from the EMNLP 2024 Paper: Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring.
|
27 |
|
28 |
+
- **Paper:** [Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring](https://arxiv.org/abs/2406.19949) (EMNLP 2024 Findings)
|
29 |
+
- **GitHub Repository:** [Thought Tree Assessment Repository](https://github.com/lijiazheng99/thought_tree_assessment)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
## Intended uses & limitations
|
32 |
|
33 |
+
This model offers a valuable resource for research in explainable AI within educational technology. The model is trained with **noisy** response-level rationales. This makes them **unsuitable** for direct application in high-stakes assessments without additional verification.
|
34 |
|
35 |
## Training and evaluation data
|
36 |
|
37 |
+
We trained and evaluated the model on the [Synthetic Rationale data](https://huggingface.co/datasets/jiazhengli/Synthetic_Rationale), which was generated from the [Rationale MCTS data](https://huggingface.co/datasets/jiazhengli/Rationale_MCTS).
|
38 |
+
|
39 |
+
To extract scores from rationales, please use the [jiazhengli/deberta-v3-large-Rationale-to-Score](https://huggingface.co/jiazhengli/deberta-v3-large-Rationale-to-Score).
|
40 |
+
|
41 |
+
## Citation
|
42 |
+
|
43 |
+
Please cite the following work if you utilize this model:
|
44 |
+
|
45 |
+
**BibTeX:**
|
46 |
+
|
47 |
+
```bibtex
|
48 |
+
@misc{li2024calibratingllmspreferenceoptimization,
|
49 |
+
title={Calibrating LLMs with Preference Optimization on Thought Trees for Generating Rationale in Science Question Scoring},
|
50 |
+
author={Jiazheng Li and Hainiu Xu and Zhaoyue Sun and Yuxiang Zhou and David West and Cesare Aloisi and Yulan He},
|
51 |
+
year={2024},
|
52 |
+
eprint={2406.19949},
|
53 |
+
archivePrefix={arXiv},
|
54 |
+
primaryClass={cs.CL},
|
55 |
+
url={https://arxiv.org/abs/2406.19949},
|
56 |
+
}
|
57 |
+
```
|
58 |
|
59 |
## Training procedure
|
60 |
|
61 |
+
Please refer to our [paper](https://arxiv.org/abs/2406.19949).
|
62 |
+
|
63 |
### Training hyperparameters
|
64 |
|
65 |
The following hyperparameters were used during training:
|