RichardErkhov
commited on
Commit
•
08969db
1
Parent(s):
b911713
uploaded readme
Browse files
README.md
ADDED
@@ -0,0 +1,430 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 |
+
LLaMAntino-3-ANITA-8B-Inst-DPO-ITA - GGUF
|
11 |
+
- Model creator: https://huggingface.co/swap-uniba/
|
12 |
+
- Original model: https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA/
|
13 |
+
|
14 |
+
|
15 |
+
| Name | Quant method | Size |
|
16 |
+
| ---- | ---- | ---- |
|
17 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q2_K.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q2_K.gguf) | Q2_K | 2.96GB |
|
18 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_XS.gguf) | IQ3_XS | 3.28GB |
|
19 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_S.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_S.gguf) | IQ3_S | 3.43GB |
|
20 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_S.gguf) | Q3_K_S | 3.41GB |
|
21 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_M.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ3_M.gguf) | IQ3_M | 3.52GB |
|
22 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K.gguf) | Q3_K | 3.74GB |
|
23 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_M.gguf) | Q3_K_M | 3.74GB |
|
24 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q3_K_L.gguf) | Q3_K_L | 4.03GB |
|
25 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ4_XS.gguf) | IQ4_XS | 4.18GB |
|
26 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_0.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_0.gguf) | Q4_0 | 4.34GB |
|
27 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.IQ4_NL.gguf) | IQ4_NL | 4.38GB |
|
28 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K_S.gguf) | Q4_K_S | 4.37GB |
|
29 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K.gguf) | Q4_K | 4.58GB |
|
30 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_K_M.gguf) | Q4_K_M | 4.58GB |
|
31 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_1.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q4_1.gguf) | Q4_1 | 4.78GB |
|
32 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_0.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_0.gguf) | Q5_0 | 5.21GB |
|
33 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K_S.gguf) | Q5_K_S | 5.21GB |
|
34 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K.gguf) | Q5_K | 5.34GB |
|
35 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_K_M.gguf) | Q5_K_M | 5.34GB |
|
36 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_1.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q5_1.gguf) | Q5_1 | 5.65GB |
|
37 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q6_K.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q6_K.gguf) | Q6_K | 6.14GB |
|
38 |
+
| [LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q8_0.gguf](https://huggingface.co/RichardErkhov/swap-uniba_-_LLaMAntino-3-ANITA-8B-Inst-DPO-ITA-gguf/blob/main/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA.Q8_0.gguf) | Q8_0 | 7.95GB |
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
Original model description:
|
44 |
+
---
|
45 |
+
language:
|
46 |
+
- en
|
47 |
+
- it
|
48 |
+
license: llama3
|
49 |
+
library_name: transformers
|
50 |
+
tags:
|
51 |
+
- facebook
|
52 |
+
- meta
|
53 |
+
- pythorch
|
54 |
+
- llama
|
55 |
+
- llama-3
|
56 |
+
- llamantino
|
57 |
+
base_model: meta-llama/Meta-Llama-3-8B-Instruct
|
58 |
+
datasets:
|
59 |
+
- gsarti/clean_mc4_it
|
60 |
+
- Chat-Error/wizard_alpaca_dolly_orca
|
61 |
+
- mlabonne/orpo-dpo-mix-40k
|
62 |
+
metrics:
|
63 |
+
- accuracy
|
64 |
+
model_creator: Marco Polignano - SWAP Research Group
|
65 |
+
pipeline_tag: text-generation
|
66 |
+
model-index:
|
67 |
+
- name: LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
68 |
+
results:
|
69 |
+
- task:
|
70 |
+
type: text-generation
|
71 |
+
name: Text Generation
|
72 |
+
dataset:
|
73 |
+
name: AI2 Reasoning Challenge (25-Shot)
|
74 |
+
type: ai2_arc
|
75 |
+
config: ARC-Challenge
|
76 |
+
split: test
|
77 |
+
args:
|
78 |
+
num_few_shot: 25
|
79 |
+
metrics:
|
80 |
+
- type: acc_norm
|
81 |
+
value: 74.57
|
82 |
+
name: normalized accuracy
|
83 |
+
source:
|
84 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
85 |
+
name: Open LLM Leaderboard
|
86 |
+
- task:
|
87 |
+
type: text-generation
|
88 |
+
name: Text Generation
|
89 |
+
dataset:
|
90 |
+
name: HellaSwag (10-Shot)
|
91 |
+
type: hellaswag
|
92 |
+
split: validation
|
93 |
+
args:
|
94 |
+
num_few_shot: 10
|
95 |
+
metrics:
|
96 |
+
- type: acc_norm
|
97 |
+
value: 92.75
|
98 |
+
name: normalized accuracy
|
99 |
+
source:
|
100 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
101 |
+
name: Open LLM Leaderboard
|
102 |
+
- task:
|
103 |
+
type: text-generation
|
104 |
+
name: Text Generation
|
105 |
+
dataset:
|
106 |
+
name: MMLU (5-Shot)
|
107 |
+
type: cais/mmlu
|
108 |
+
config: all
|
109 |
+
split: test
|
110 |
+
args:
|
111 |
+
num_few_shot: 5
|
112 |
+
metrics:
|
113 |
+
- type: acc
|
114 |
+
value: 66.85
|
115 |
+
name: accuracy
|
116 |
+
source:
|
117 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
118 |
+
name: Open LLM Leaderboard
|
119 |
+
- task:
|
120 |
+
type: text-generation
|
121 |
+
name: Text Generation
|
122 |
+
dataset:
|
123 |
+
name: TruthfulQA (0-shot)
|
124 |
+
type: truthful_qa
|
125 |
+
config: multiple_choice
|
126 |
+
split: validation
|
127 |
+
args:
|
128 |
+
num_few_shot: 0
|
129 |
+
metrics:
|
130 |
+
- type: mc2
|
131 |
+
value: 75.93
|
132 |
+
source:
|
133 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
134 |
+
name: Open LLM Leaderboard
|
135 |
+
- task:
|
136 |
+
type: text-generation
|
137 |
+
name: Text Generation
|
138 |
+
dataset:
|
139 |
+
name: Winogrande (5-shot)
|
140 |
+
type: winogrande
|
141 |
+
config: winogrande_xl
|
142 |
+
split: validation
|
143 |
+
args:
|
144 |
+
num_few_shot: 5
|
145 |
+
metrics:
|
146 |
+
- type: acc
|
147 |
+
value: 82.0
|
148 |
+
name: accuracy
|
149 |
+
source:
|
150 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
151 |
+
name: Open LLM Leaderboard
|
152 |
+
- task:
|
153 |
+
type: text-generation
|
154 |
+
name: Text Generation
|
155 |
+
dataset:
|
156 |
+
name: GSM8k (5-shot)
|
157 |
+
type: gsm8k
|
158 |
+
config: main
|
159 |
+
split: test
|
160 |
+
args:
|
161 |
+
num_few_shot: 5
|
162 |
+
metrics:
|
163 |
+
- type: acc
|
164 |
+
value: 58.61
|
165 |
+
name: accuracy
|
166 |
+
source:
|
167 |
+
url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA
|
168 |
+
name: Open LLM Leaderboard
|
169 |
+
---
|
170 |
+
<img src="https://cdn-uploads.huggingface.co/production/uploads/5df8bb21da6d0311fd3d540f/cZoZdwQOPdQsnQmDXHcSn.png" alt="llamantino3_anita" border="0" width="800px">
|
171 |
+
|
172 |
+
<hr>
|
173 |
+
<!--<img src="https://i.ibb.co/6mHSRm3/llamantino53.jpg" width="200"/>-->
|
174 |
+
<h3><i>"Built with <b>Meta Llama 3</b>".</i></i></h3>
|
175 |
+
<p style="text-align:justify;"><b>LLaMAntino-3-ANITA-8B-Inst-DPO-ITA</b> is a model of the <a href="https://huggingface.co/swap-uniba"><b>LLaMAntino</b></a> - <i>Large Language Models family</i>.
|
176 |
+
The model is an instruction-tuned version of <a href="https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct"><b>Meta-Llama-3-8b-instruct</b></a> (a fine-tuned <b>LLaMA 3 model</b>).
|
177 |
+
This model version aims to be the a <b>Multilingual Model</b> 🏁 (EN 🇺🇸 + ITA🇮🇹) to further fine-tuning on Specific Tasks in Italian.</p>
|
178 |
+
|
179 |
+
|
180 |
+
The 🌟**ANITA project**🌟 *(**A**dvanced **N**atural-based interaction for the **ITA**lian language)*
|
181 |
+
wants to provide Italian NLP researchers with an improved model for the Italian Language 🇮🇹 use cases.<br>
|
182 |
+
|
183 |
+
<hr>
|
184 |
+
|
185 |
+
**Live DEMO:** [https://chat.llamantino.it/](https://chat.llamantino.it/)<br>
|
186 |
+
*It works only with Italian connection.*
|
187 |
+
|
188 |
+
<hr>
|
189 |
+
|
190 |
+
## Model Details
|
191 |
+
*Last Update: 10/05/2024*<br>
|
192 |
+
|
193 |
+
<a href="https://github.com/marcopoli/LLaMAntino-3-ANITA"><img src="https://github.githubassets.com/assets/GitHub-Logo-ee398b662d42.png" width="150"> https://github.com/marcopoli/LLaMAntino-3-ANITA</a><br>
|
194 |
+
|
195 |
+
| Model | HF | GGUF | EXL2 |
|
196 |
+
|-------|-------|-------|-------|
|
197 |
+
| *swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA* | [Link](https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA) | [Link](https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA_GGUF) | [Link](https://huggingface.co/swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA_EXL2) |
|
198 |
+
|
199 |
+
<hr>
|
200 |
+
|
201 |
+
## Specifications
|
202 |
+
|
203 |
+
- **Model developers**: <br><a href="https://marcopoli.github.io/">Ph.D. Marco Polignano</a> - University of Bari Aldo Moro, Italy <br> <a href="https://huggingface.co/swap-uniba">SWAP Research Group</a> <br>
|
204 |
+
- **Variations**: The model release has been **supervised fine-tuning (SFT)** using **QLoRA** 4bit, on instruction-based datasets. **DPO** approach over the *mlabonne/orpo-dpo-mix-40k* dataset is used to align with human preferences for helpfulness and safety.
|
205 |
+
- **Input**: Models input text only.
|
206 |
+
- **Language**: Multilingual 🏁 + Italian 🇮🇹
|
207 |
+
- **Output**: Models generate text and code only.
|
208 |
+
- **Model Architecture**: *Llama 3 architecture*.
|
209 |
+
- **Context length**: 8K, 8192.
|
210 |
+
- **Library Used**: [Unsloth](https://unsloth.ai/)
|
211 |
+
<hr>
|
212 |
+
|
213 |
+
## Playground
|
214 |
+
|
215 |
+
To use the model directly, there are many ways to get started, choose one of the following ways to experience it.
|
216 |
+
|
217 |
+
### Prompt Template
|
218 |
+
```
|
219 |
+
<|start_header_id|>system<|end_header_id|>
|
220 |
+
|
221 |
+
{ SYS Prompt }<|eot_id|><|start_header_id|>user<|end_header_id|>
|
222 |
+
|
223 |
+
{ USER Prompt }<|eot_id|><|start_header_id|>assistant<|end_header_id|>
|
224 |
+
|
225 |
+
{ ASSIST Prompt }<|eot_id|>
|
226 |
+
````
|
227 |
+
|
228 |
+
### Transformers
|
229 |
+
|
230 |
+
For direct use with `transformers`, you can easily get started with the following steps.
|
231 |
+
|
232 |
+
- Firstly, you need to install transformers via the command below with `pip`.
|
233 |
+
|
234 |
+
```bash
|
235 |
+
pip install -U transformers trl peft accelerate bitsandbytes
|
236 |
+
```
|
237 |
+
|
238 |
+
- Right now, you can start using the model directly.
|
239 |
+
|
240 |
+
```python
|
241 |
+
import torch
|
242 |
+
from transformers import (
|
243 |
+
AutoModelForCausalLM,
|
244 |
+
AutoTokenizer,
|
245 |
+
)
|
246 |
+
|
247 |
+
base_model = "swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA"
|
248 |
+
model = AutoModelForCausalLM.from_pretrained(
|
249 |
+
base_model,
|
250 |
+
torch_dtype=torch.bfloat16,
|
251 |
+
device_map="auto",
|
252 |
+
)
|
253 |
+
tokenizer = AutoTokenizer.from_pretrained(base_model)
|
254 |
+
|
255 |
+
sys = "Sei un an assistente AI per la lingua Italiana di nome LLaMAntino-3 ANITA " \
|
256 |
+
"(Advanced Natural-based interaction for the ITAlian language)." \
|
257 |
+
" Rispondi nella lingua usata per la domanda in modo chiaro, semplice ed esaustivo."
|
258 |
+
|
259 |
+
messages = [
|
260 |
+
{"role": "system", "content": sys},
|
261 |
+
{"role": "user", "content": "Chi è Carlo Magno?"}
|
262 |
+
]
|
263 |
+
|
264 |
+
#Method 1
|
265 |
+
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
266 |
+
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
|
267 |
+
for k,v in inputs.items():
|
268 |
+
inputs[k] = v.cuda()
|
269 |
+
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, top_p=0.9, temperature=0.6)
|
270 |
+
results = tokenizer.batch_decode(outputs)[0]
|
271 |
+
print(results)
|
272 |
+
|
273 |
+
#Method 2
|
274 |
+
import transformers
|
275 |
+
pipe = transformers.pipeline(
|
276 |
+
model=model,
|
277 |
+
tokenizer=tokenizer,
|
278 |
+
return_full_text=False, # langchain expects the full text
|
279 |
+
task='text-generation',
|
280 |
+
max_new_tokens=512, # max number of tokens to generate in the output
|
281 |
+
temperature=0.6, #temperature for more or less creative answers
|
282 |
+
do_sample=True,
|
283 |
+
top_p=0.9,
|
284 |
+
)
|
285 |
+
|
286 |
+
sequences = pipe(messages)
|
287 |
+
for seq in sequences:
|
288 |
+
print(f"{seq['generated_text']}")
|
289 |
+
|
290 |
+
```
|
291 |
+
|
292 |
+
- Additionally, you can also use a model with **4bit quantization** to reduce the required resources at least. You can start with the code below.
|
293 |
+
|
294 |
+
```python
|
295 |
+
import torch
|
296 |
+
from transformers import (
|
297 |
+
AutoModelForCausalLM,
|
298 |
+
AutoTokenizer,
|
299 |
+
BitsAndBytesConfig,
|
300 |
+
)
|
301 |
+
|
302 |
+
base_model = "swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA"
|
303 |
+
bnb_config = BitsAndBytesConfig(
|
304 |
+
load_in_4bit=True,
|
305 |
+
bnb_4bit_quant_type="nf4",
|
306 |
+
bnb_4bit_compute_dtype=torch.bfloat16,
|
307 |
+
bnb_4bit_use_double_quant=False,
|
308 |
+
)
|
309 |
+
model = AutoModelForCausalLM.from_pretrained(
|
310 |
+
base_model,
|
311 |
+
quantization_config=bnb_config,
|
312 |
+
device_map="auto",
|
313 |
+
)
|
314 |
+
tokenizer = AutoTokenizer.from_pretrained(base_model)
|
315 |
+
|
316 |
+
sys = "Sei un an assistente AI per la lingua Italiana di nome LLaMAntino-3 ANITA " \
|
317 |
+
"(Advanced Natural-based interaction for the ITAlian language)." \
|
318 |
+
" Rispondi nella lingua usata per la domanda in modo chiaro, semplice ed esaustivo."
|
319 |
+
|
320 |
+
messages = [
|
321 |
+
{"role": "system", "content": sys},
|
322 |
+
{"role": "user", "content": "Chi è Carlo Magno?"}
|
323 |
+
]
|
324 |
+
|
325 |
+
#Method 1
|
326 |
+
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
327 |
+
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
|
328 |
+
for k,v in inputs.items():
|
329 |
+
inputs[k] = v.cuda()
|
330 |
+
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, top_p=0.9, temperature=0.6)
|
331 |
+
results = tokenizer.batch_decode(outputs)[0]
|
332 |
+
print(results)
|
333 |
+
|
334 |
+
#Method 2
|
335 |
+
import transformers
|
336 |
+
pipe = transformers.pipeline(
|
337 |
+
model=model,
|
338 |
+
tokenizer=tokenizer,
|
339 |
+
return_full_text=False, # langchain expects the full text
|
340 |
+
task='text-generation',
|
341 |
+
max_new_tokens=512, # max number of tokens to generate in the output
|
342 |
+
temperature=0.6, #temperature for more or less creative answers
|
343 |
+
do_sample=True,
|
344 |
+
top_p=0.9,
|
345 |
+
)
|
346 |
+
|
347 |
+
sequences = pipe(messages)
|
348 |
+
for seq in sequences:
|
349 |
+
print(f"{seq['generated_text']}")
|
350 |
+
|
351 |
+
```
|
352 |
+
|
353 |
+
<hr>
|
354 |
+
|
355 |
+
## Evaluation
|
356 |
+
|
357 |
+
**Open LLM Leaderboard:**
|
358 |
+
|
359 |
+
Evaluated with lm-evaluation-benchmark-harness for the [**Open Italian LLMs Leaderboard**](https://huggingface.co/spaces/FinancialSupport/open_ita_llm_leaderboard)
|
360 |
+
```
|
361 |
+
lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID --tasks hellaswag_it,arc_it --device cuda:0 --batch_size auto:2
|
362 |
+
lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID --tasks m_mmlu_it --num_fewshot 5 --device cuda:0 --batch_size auto:2
|
363 |
+
```
|
364 |
+
|
365 |
+
| Metric | Value |
|
366 |
+
|-----------------------|---------------------------|
|
367 |
+
| Avg. | **0.6160** |
|
368 |
+
| Arc_IT | 0.5714 |
|
369 |
+
| Hellaswag_IT | 0.7093 |
|
370 |
+
| MMLU_IT | 0.5672 |
|
371 |
+
|
372 |
+
<hr>
|
373 |
+
|
374 |
+
## Unsloth
|
375 |
+
|
376 |
+
<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/made with unsloth.png" width="200px" align="center" />
|
377 |
+
|
378 |
+
[Unsloth](https://unsloth.ai), a great tool that helps us easily develop products, at a lower cost than expected.
|
379 |
+
|
380 |
+
## Citation instructions
|
381 |
+
```bibtex
|
382 |
+
@misc{polignano2024advanced,
|
383 |
+
title={Advanced Natural-based interaction for the ITAlian language: LLaMAntino-3-ANITA},
|
384 |
+
author={Marco Polignano and Pierpaolo Basile and Giovanni Semeraro},
|
385 |
+
year={2024},
|
386 |
+
eprint={2405.07101},
|
387 |
+
archivePrefix={arXiv},
|
388 |
+
primaryClass={cs.CL}
|
389 |
+
}
|
390 |
+
```
|
391 |
+
|
392 |
+
```bibtex
|
393 |
+
@misc{basile2023llamantino,
|
394 |
+
title={LLaMAntino: LLaMA 2 Models for Effective Text Generation in Italian Language},
|
395 |
+
author={Pierpaolo Basile and Elio Musacchio and Marco Polignano and Lucia Siciliani and Giuseppe Fiameni and Giovanni Semeraro},
|
396 |
+
year={2023},
|
397 |
+
eprint={2312.09993},
|
398 |
+
archivePrefix={arXiv},
|
399 |
+
primaryClass={cs.CL}
|
400 |
+
}
|
401 |
+
```
|
402 |
+
|
403 |
+
```bibtex
|
404 |
+
@article{llama3modelcard,
|
405 |
+
title={Llama 3 Model Card},
|
406 |
+
author={AI@Meta},
|
407 |
+
year={2024},
|
408 |
+
url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md}
|
409 |
+
}
|
410 |
+
```
|
411 |
+
|
412 |
+
# Acknowledgments
|
413 |
+
We acknowledge the support of the PNRR project [FAIR - Future AI Research (PE00000013)](https://fondazione-fair.it/en/foundation/), Spoke 6 - Symbiotic AI (CUP H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU.
|
414 |
+
Models are built on the Leonardo supercomputer with the support of CINECA-Italian Super Computing Resource Allocation, class C project IscrC\_Pro\_MRS (HP10CQO70G).
|
415 |
+
<img src="https://wiki.u-gov.it/confluence/download/attachments/49842317/image2022-6-21_11-11-44.png?version=1&modificationDate=1655802705000&api=v2" width="600px">
|
416 |
+
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
417 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_swap-uniba__LLaMAntino-3-ANITA-8B-Inst-DPO-ITA)
|
418 |
+
|
419 |
+
| Metric |Value|
|
420 |
+
|---------------------------------|----:|
|
421 |
+
|Avg. |75.12|
|
422 |
+
|AI2 Reasoning Challenge (25-Shot)|74.57|
|
423 |
+
|HellaSwag (10-Shot) |92.75|
|
424 |
+
|MMLU (5-Shot) |66.85|
|
425 |
+
|TruthfulQA (0-shot) |75.93|
|
426 |
+
|Winogrande (5-shot) |82.00|
|
427 |
+
|GSM8k (5-shot) |58.61|
|
428 |
+
|
429 |
+
|
430 |
+
|