Hugging Face's logo

RichardErkhov
/
PKU-ONELab_-_Themis-gguf

GGUF
Model card Files Community
RichardErkhov commited on
Commit
33597c1
·
verified ·
1 Parent(s): b47078d

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +109 -0
README.md ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ Themis - GGUF
11
+ - Model creator: https://huggingface.co/PKU-ONELab/
12
+ - Original model: https://huggingface.co/PKU-ONELab/Themis/
13
+
14
+
15
+ | Name | Quant method | Size |
16
+ | ---- | ---- | ---- |
17
+ | [Themis.Q2_K.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q2_K.gguf) | Q2_K | 2.96GB |
18
+ | [Themis.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q3_K_S.gguf) | Q3_K_S | 3.41GB |
19
+ | [Themis.Q3_K.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q3_K.gguf) | Q3_K | 3.74GB |
20
+ | [Themis.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q3_K_M.gguf) | Q3_K_M | 3.74GB |
21
+ | [Themis.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q3_K_L.gguf) | Q3_K_L | 4.03GB |
22
+ | [Themis.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.IQ4_XS.gguf) | IQ4_XS | 4.18GB |
23
+ | [Themis.Q4_0.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q4_0.gguf) | Q4_0 | 4.34GB |
24
+ | [Themis.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.IQ4_NL.gguf) | IQ4_NL | 4.38GB |
25
+ | [Themis.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q4_K_S.gguf) | Q4_K_S | 4.37GB |
26
+ | [Themis.Q4_K.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q4_K.gguf) | Q4_K | 4.58GB |
27
+ | [Themis.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q4_K_M.gguf) | Q4_K_M | 4.58GB |
28
+ | [Themis.Q4_1.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q4_1.gguf) | Q4_1 | 4.78GB |
29
+ | [Themis.Q5_0.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q5_0.gguf) | Q5_0 | 5.21GB |
30
+ | [Themis.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q5_K_S.gguf) | Q5_K_S | 5.21GB |
31
+ | [Themis.Q5_K.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q5_K.gguf) | Q5_K | 5.34GB |
32
+ | [Themis.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q5_K_M.gguf) | Q5_K_M | 5.34GB |
33
+ | [Themis.Q5_1.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q5_1.gguf) | Q5_1 | 5.65GB |
34
+ | [Themis.Q6_K.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q6_K.gguf) | Q6_K | 6.14GB |
35
+ | [Themis.Q8_0.gguf](https://huggingface.co/RichardErkhov/PKU-ONELab_-_Themis-gguf/blob/main/Themis.Q8_0.gguf) | Q8_0 | 7.95GB |
36
+
37
+
38
+
39
+
40
+ Original model description:
41
+ ---
42
+ license: apache-2.0
43
+ ---
44
+ # Themis
45
+
46
+ Paper: https://arxiv.org/abs/2406.18365
47
+
48
+ Github: https://github.com/PKU-ONELab/Themis
49
+
50
+ ## Introduction
51
+
52
+ We propose **Themis**, an 8B-parameter large language model (LLM) specifically designed and trained for NLG evaluation with more comprehensive capabilities.
53
+
54
+ Our Themis can evaluate various NLG tasks, including uncommon ones like question-answering evaluation (**Versatility**), in a reference-free manner (**Independence**). Moreover, it allows for specific and customized evaluation aspects and criteria, including overall quality and more fine-grained aspects (**Flexibility**), and its evaluation contains corresponding analysis and explanation together with the rating (**Interpretability**).
55
+
56
+ We believe that an ideal evaluator should be convenient to use and possess these characteristics. The comparison between related methods and Themis is shown in the table below.
57
+
58
+ | Method | Versatility | Independence | Flexibility | Interpretability | Open-source |
59
+ | :---------------: | :---------: | :----------: | :---------: | :--------------: | :---------: |
60
+ | UniEval | ❌ | ❌ | ✔️ | ❌ | ✔️ |
61
+ | G-Eval | ✔️ | ✔️ | ✔️ | ✔️ | ❌ |
62
+ | X-Eval | ✔️ | ❌ | ✔️ | ❌ | ❌ |
63
+ | Prometheus | ✔️ | ❌ | ✔️ | ✔️ | ✔️ |
64
+ | Auto-J | ✔️ | ✔️ | ❌ | ✔️ | ✔️ |
65
+ | InstructScore | ✔️ | ❌ | ❌ | ✔️ | ✔️ |
66
+ | TIGERScore | ✔️ | ✔️ | ❌ | ✔️ | ✔️ |
67
+ | **Themis (Ours)** | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
68
+
69
+ ## Performance
70
+
71
+ We implement experiments on several common NLG evaluation tasks and datasets to compare our Themis with other methods, including SummEval for summarization, Topical-Chat for dialogue response generation, SFRES&SFHOT for data-to-text, QAGS for factuality, MANS for story generation, and WMT23 zh-en for machine translation. Experimental results show that our Themis achieves better overall evaluation performance over other evaluation models, including GPT-4.
72
+
73
+ | Method | SummEval | Topical-Chat | SFHOT& SFRES | QAGS | MANS | WMT23 | Average Spearman |
74
+ | -------------------- | :-------: | :----------: | :---------: | :-------: | :-------: | :-------: | :------------: |
75
+ | BLEU | 0.075 | 0.388 | 0.024 | - | 0.032 | 0.021 | - |
76
+ | ROUGE | 0.152 | 0.412 | 0.101 | - | -0.002 | 0.151 | - |
77
+ | BARTScore | 0.329 | 0.086 | 0.208 | 0.425 | 0.350 | 0.118 | 0.253 |
78
+ | BERTScore | 0.231 | 0.394 | 0.139 | - | 0.285 | 0.219 | - |
79
+ | BLEURT | 0.152 | 0.388 | 0.244 | - | 0.138 | 0.263 | - |
80
+ | CometKiwi | 0.228 | 0.340 | 0.251 | 0.094 | 0.251 | 0.343 | 0.251 |
81
+ | UniEval | 0.474 | 0.577 | 0.282 | - | - | - | - |
82
+ | G-Eval (GPT-3.5) | 0.409 | 0.585 | - | 0.461 | - | - | - |
83
+ | G-Eval (GPT-4) | 0.523 | 0.588 | - | 0.611 | - | - | - |
84
+ | GPT-3.5 Turbo | 0.416 | 0.578 | 0.306 | 0.431 | 0.328 | 0.347 | 0.401 |
85
+ | GPT-4 Turbo | 0.511 | **0.746** | 0.320 | 0.637 | 0.473 | **0.437** | 0.521 |
86
+ | X-Eval | 0.480 | 0.605 | 0.303 | 0.578 | - | - | - |
87
+ | Prometheus-13B | 0.163 | 0.434 | 0.173 | - | 0.007 | 0.129 | - |
88
+ | Auto-J-13B | 0.198 | 0.425 | 0.141 | 0.226 | 0.380 | 0.104 | 0.246 |
89
+ | TIGERScore-13B | 0.384 | 0.346 | 0.200 | 0.504 | 0.231 | 0.248 | 0.319 |
90
+ | InstructScore-7B | 0.258 | 0.241 | 0.247 | - | 0.298 | 0.219 | - |
91
+ | **Themis-8B (ours)** | **0.553** | 0.725 | **0.333** | **0.684** | **0.551** | 0.405 | **0.542** |
92
+
93
+ We further conduct more in-depth analyses, including generalization tests on unseen tasks like the instruction-following evaluation as well as aspect-targeted perturbation tests, and our Themis also exhibits superior evaluation performance. For more experimental results and details, please refer to our paper.
94
+
95
+ ## Requirements and Usage
96
+
97
+ Please refer to our [github repo](https://github.com/PKU-ONELab/Themis) for more details.
98
+
99
+ ## Citation
100
+
101
+ ```
102
+ @article{hu2024themis,
103
+ title={Themis: Towards Flexible and Interpretable NLG Evaluation},
104
+ author={Hu, Xinyu and Lin, Li and Gao, Mingqi and Yin, Xunjian and Wan, Xiaojun},
105
+ journal={arXiv preprint arXiv:2406.18365},
106
+ year={2024}
107
+ }
108
+ ```
109
+