Hugging Face's logo

mlabonne
/
Marcoro14-7B-slerp

Text Generation
Transformers
Safetensors
mistral
Merge
mergekit
lazymergekit
AIDC-ai-business/Marcoroni-7B-v3
EmbeddedLLM/Mistral-7B-Merge-14-v0.1
Eval Results
text-generation-inference
Inference Endpoints
Model card Files Community
2
mlabonne commited on
Commit
38aba7d
•
1 Parent(s): 777b2c7

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +113 -1
README.md CHANGED
@@ -5,12 +5,99 @@ tags:
5
  - mergekit
6
  ---
7
 
 
 
8
  # Marcoro14-7B-slerp
9
 
10
  This model is a merge of the following models made with [mergekit](https://github.com/cg123/mergekit):
11
  * [AIDC-ai-business/Marcoroni-7B-v3](https://huggingface.co/AIDC-ai-business/Marcoroni-7B-v3)
12
  * [EmbeddedLLM/Mistral-7B-Merge-14-v0.1](https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1)
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ## 🧩 Configuration
15
 
16
  ```yaml
@@ -30,4 +117,29 @@ parameters:
30
  value: [1, 0.5, 0.7, 0.3, 0]
31
  - value: 0.5
32
  dtype: bfloat16
33
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  - mergekit
6
  ---
7
 
8
+ ![](https://i.imgur.com/FSKtmRc.png)
9
+
10
  # Marcoro14-7B-slerp
11
 
12
  This model is a merge of the following models made with [mergekit](https://github.com/cg123/mergekit):
13
  * [AIDC-ai-business/Marcoroni-7B-v3](https://huggingface.co/AIDC-ai-business/Marcoroni-7B-v3)
14
  * [EmbeddedLLM/Mistral-7B-Merge-14-v0.1](https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0.1)
15
 
16
+ ## 🏆 Evaluation
17
+
18
+ Marcoro14-7B-slerp is the second best-performing 7B LLM on the Open LLM Leaderboard:
19
+
20
+ ![](https://i.imgur.com/5XUuP7g.png)
21
+
22
+ I also evaluated it using Nous' benchmark suite and obtained the following results:
23
+
24
+ | Model |agieval|gpt4all|truthfulqa|bigbench|Average|
25
+ |-------------------------|------:|------:|---------:|-------:|------:|
26
+ |Marcoro14-7B-slerp | 44.66| 76.24| 64.15| 45.64| 57.67|
27
+ |OpenHermes-2.5-Mistral-7B| 43.07| 73.12| 53.04| 40.96| 52.57|
28
+ |Change | +1.59| +3.12| +11.11| +4.68| +5.1|
29
+
30
+ ### AGIEVAL
31
+ | Task |Version| Metric |Value| |Stderr|
32
+ |------------------------------|------:|--------|----:|---|-----:|
33
+ |agieval_aqua_rat | 0|acc |26.38|± | 2.77|
34
+ | | |acc_norm|24.41|± | 2.70|
35
+ |agieval_logiqa_en | 0|acc |38.25|± | 1.91|
36
+ | | |acc_norm|39.32|± | 1.92|
37
+ |agieval_lsat_ar | 0|acc |24.35|± | 2.84|
38
+ | | |acc_norm|25.22|± | 2.87|
39
+ |agieval_lsat_lr | 0|acc |50.00|± | 2.22|
40
+ | | |acc_norm|50.59|± | 2.22|
41
+ |agieval_lsat_rc | 0|acc |62.83|± | 2.95|
42
+ | | |acc_norm|62.08|± | 2.96|
43
+ |agieval_sat_en | 0|acc |79.61|± | 2.81|
44
+ | | |acc_norm|79.61|± | 2.81|
45
+ |agieval_sat_en_without_passage| 0|acc |45.15|± | 3.48|
46
+ | | |acc_norm|45.63|± | 3.48|
47
+ |agieval_sat_math | 0|acc |33.18|± | 3.18|
48
+ | | |acc_norm|30.45|± | 3.11|
49
+ Average: 44.66%
50
+
51
+ ### GPT4ALL
52
+ | Task |Version| Metric |Value| |Stderr|
53
+ |-------------|------:|--------|----:|---|-----:|
54
+ |arc_challenge| 0|acc |63.91|± | 1.40|
55
+ | | |acc_norm|64.93|± | 1.39|
56
+ |arc_easy | 0|acc |86.07|± | 0.71|
57
+ | | |acc_norm|83.75|± | 0.76|
58
+ |boolq | 1|acc |88.56|± | 0.56|
59
+ |hellaswag | 0|acc |67.31|± | 0.47|
60
+ | | |acc_norm|85.28|± | 0.35|
61
+ |openbookqa | 0|acc |36.40|± | 2.15|
62
+ | | |acc_norm|48.20|± | 2.24|
63
+ |piqa | 0|acc |82.59|± | 0.88|
64
+ | | |acc_norm|84.39|± | 0.85|
65
+ |winogrande | 0|acc |78.53|± | 1.15|
66
+ Average: 76.24%
67
+
68
+ ### TRUTHFULQA
69
+ | Task |Version|Metric|Value| |Stderr|
70
+ |-------------|------:|------|----:|---|-----:|
71
+ |truthfulqa_mc| 1|mc1 |46.88|± | 1.75|
72
+ | | |mc2 |64.15|± | 1.52|
73
+ Average: 64.15%
74
+
75
+ ### BIGBENCH
76
+ | Task |Version| Metric |Value| |Stderr|
77
+ |------------------------------------------------|------:|---------------------|----:|---|-----:|
78
+ |bigbench_causal_judgement | 0|multiple_choice_grade|56.32|± | 3.61|
79
+ |bigbench_date_understanding | 0|multiple_choice_grade|66.40|± | 2.46|
80
+ |bigbench_disambiguation_qa | 0|multiple_choice_grade|45.35|± | 3.11|
81
+ |bigbench_geometric_shapes | 0|multiple_choice_grade|20.33|± | 2.13|
82
+ | | |exact_str_match | 4.74|± | 1.12|
83
+ |bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|30.00|± | 2.05|
84
+ |bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|21.43|± | 1.55|
85
+ |bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|52.33|± | 2.89|
86
+ |bigbench_movie_recommendation | 0|multiple_choice_grade|39.20|± | 2.19|
87
+ |bigbench_navigate | 0|multiple_choice_grade|53.90|± | 1.58|
88
+ |bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|72.15|± | 1.00|
89
+ |bigbench_ruin_names | 0|multiple_choice_grade|52.46|± | 2.36|
90
+ |bigbench_salient_translation_error_detection | 0|multiple_choice_grade|25.75|± | 1.38|
91
+ |bigbench_snarks | 0|multiple_choice_grade|72.38|± | 3.33|
92
+ |bigbench_sports_understanding | 0|multiple_choice_grade|73.63|± | 1.40|
93
+ |bigbench_temporal_sequences | 0|multiple_choice_grade|45.70|± | 1.58|
94
+ |bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|23.44|± | 1.20|
95
+ |bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|18.51|± | 0.93|
96
+ |bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|52.33|± | 2.89|
97
+ Average: 45.64%
98
+
99
+ Average score: 57.67%
100
+
101
  ## 🧩 Configuration
102
 
103
  ```yaml
 
117
  value: [1, 0.5, 0.7, 0.3, 0]
118
  - value: 0.5
119
  dtype: bfloat16
120
+ ```
121
+
122
+ ## 💻 Usage
123
+
124
+ ```python
125
+ !pip install -qU transformers accelerate
126
+
127
+ from transformers import AutoTokenizer
128
+ import transformers
129
+ import torch
130
+
131
+ model = "mlabonne/NeuralPipe-7B-slerp"
132
+ messages = [{"role": "user", "content": "What is a large language model?"}]
133
+
134
+ tokenizer = AutoTokenizer.from_pretrained(model)
135
+ prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
136
+ pipeline = transformers.pipeline(
137
+ "text-generation",
138
+ model=model,
139
+ torch_dtype=torch.float16,
140
+ device_map="auto",
141
+ )
142
+
143
+ outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
144
+ print(outputs[0]["generated_text"])
145
+ ```