File size: 4,030 Bytes
9fba537
 
da94cac
4f0bf31
9fba537
 
410f08f
9fba537
 
4f0bf31
9fba537
 
 
 
 
 
 
4f0bf31
 
 
 
 
9fba537
 
 
 
2525d11
9fba537
e209d7d
9fba537
ad2b123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9fba537
e209d7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9fba537
e209d7d
9fba537
4f0bf31
 
9fba537
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
library_name: peft
base_model: NousResearch/Meta-Llama-3-70B-Instruct
license: apache-2.0
---

# Model Card for radm/Llama-3-70B-Instruct-AH-lora

<!-- Provide a quick summary of what the model is/does. -->
This is a LORA adapter for NousResearch/Meta-Llama-3-70B-Instruct, fine-tuned to be a judge on Arena Hard (https://github.com/lm-sys/arena-hard-auto)


## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->
- **Developed by:** [radm]
- **Model type:** [Llama-3-70b]
- **Language(s) (NLP):** [English]
- **License:** [apache-2.0]
- **Finetuned from model [optional]:** [NousResearch/Meta-Llama-3-70B-Instruct]

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
Use repository (https://github.com/r4dm/arena-hard-local) for evaluate with local judge model.

## Results

#### Llama-3-70B-Instruct-GPTQ as judge:
```console
Llama-3-Instruct-8B-SimPO                          | score: 78.3  | 95% CI:   (-1.5, 1.2)   | average #tokens: 545
SELM-Llama-3-8B-Instruct-iter-3                    | score: 72.8  | 95% CI:   (-2.1, 1.4)   | average #tokens: 606
Meta-Llama-3-8B-Instruct-f16                       | score: 65.3  | 95% CI:   (-1.8, 2.1)   | average #tokens: 560
suzume-llama-3-8B-multilingual-orpo-borda-half     | score: 63.5  | 95% CI:   (-1.6, 2.1)   | average #tokens: 978
Phi-3-medium-128k-instruct                         | score: 50.0  | 95% CI:   (0.0, 0.0)    | average #tokens: 801
suzume-llama-3-8B-multilingual                     | score: 48.1  | 95% CI:   (-2.2, 1.8)   | average #tokens: 767
aya-23-8B                                          | score: 48.0  | 95% CI:   (-2.0, 2.1)   | average #tokens: 834
Vikhr-7B-instruct_0.5                              | score: 19.6  | 95% CI:   (-1.3, 1.5)   | average #tokens: 794
alpindale_gemma-2b-it                              | score: 11.2  | 95% CI:   (-1.0, 0.8)   | average #tokens: 425
```
#### Llama-3-70B-Instruct-AH-AWQ as judge:
```console
Llama-3-Instruct-8B-SimPO                          | score: 83.8  | 95% CI:   (-1.4, 1.3)   | average #tokens: 545
SELM-Llama-3-8B-Instruct-iter-3                    | score: 78.8  | 95% CI:   (-1.7, 1.9)   | average #tokens: 606
suzume-llama-3-8B-multilingual-orpo-borda-half     | score: 71.8  | 95% CI:   (-1.7, 2.4)   | average #tokens: 978
Meta-Llama-3-8B-Instruct-f16                       | score: 69.8  | 95% CI:   (-1.9, 1.7)   | average #tokens: 560
suzume-llama-3-8B-multilingual                     | score: 54.0  | 95% CI:   (-2.1, 2.1)   | average #tokens: 767
aya-23-8B                                          | score: 50.4  | 95% CI:   (-1.7, 1.7)   | average #tokens: 834
Phi-3-medium-128k-instruct                         | score: 50.0  | 95% CI:   (0.0, 0.0)    | average #tokens: 801
Vikhr-7B-instruct_0.5                              | score: 14.2  | 95% CI:   (-1.3, 1.0)   | average #tokens: 794
alpindale_gemma-2b-it                              | score:  7.9  | 95% CI:   (-0.9, 0.8)   | average #tokens: 425
```

## Training Details

### Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
Datasets:
- radm/arenahard_gpt4vsllama3
- radm/truthy-dpo-v0.1-ru
- jondurbin/truthy-dpo-v0.1

#### Training Hyperparameters

- **Training regime:** [bf16] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
- **Load in 4 bit:** [True]
- **Target modules:** [all]
- **LoRA rank:** [16]
- **Max seq length:** [8192]
- **Use gradient checkpointing:** [unsloth]
- **trainer:** [ORPOTrainer]
- **Batch size:** [1]
- **Gradient accumulation steps:** [4]
- **Epochs:** [1]

### Hardware

- **Hardware Type:** [Nvidia A100 80 gb]
- **Hours used:** [11 hours]

### Framework versions

- PEFT 0.10.0