README.md

metadata

language:
  - pt
  - en
license: cc
tags:
  - text-generation-inference
  - transformers
  - mistral
  - gguf
  - brazil
  - brasil
  - portuguese
base_model: mistralai/Mistral-7B-Instruct-v0.2
pipeline_tag: text-generation
metrics:
  - name: assin2_rte f1_macro
    type: assin2_rte
    value: 90.13
  - name: assin2_rte acc
    type: assin2_rte
    value: 90.16
  - name: assin2_sts pearson
    type: assin2_sts
    value: 71.51
  - name: assin2_sts mse
    type: assin2_sts
    value: 68.03
  - name: bluex acc
    type: bluex
    value: 47.98
  - name: enem acc
    type: enem
    value: 58.43
  - name: faquad_nli f1_macro
    type: faquad_nli
    value: 64.24
  - name: faquad_nli acc
    type: faquad_nli
    value: 67.69
  - name: hatebr_offensive_binary f1_macro
    type: hatebr_offensive_binary
    value: 83.61
  - name: hatebr_offensive_binary acc
    type: hatebr_offensive_binary
    value: 83.71
  - name: oab_exams acc
    type: oab_exams
    value: 38.41
  - name: portuguese_hate_speech_binary f1_macro
    type: portuguese_hate_speech_binary
    value: 61.87
  - name: portuguese_hate_speech_binary acc
    type: portuguese_hate_speech_binary
    value: 63.22

Cabra Mistral 7b v2

Esse modelo é um finetune do Mistral 7b Instruct 0.2 com o dataset interno Cabra 10k. Esse modelo é optimizado para português e responde em portuguese nativamente. Ele apresenta melhoria em varios benchmarks brasileiros em comparação com o modelo base.

Exprimente o nosso demo aqui: CabraChat.

Conheça os nossos outros modelos: Cabra.

Detalhes do Modelo

Modelo: Mistral 7b Instruct 0.2

Mistral-7B-v0.1 é um modelo de transformador, com as seguintes escolhas arquitetônicas:

• Grouped-Query Attention
• Sliding-Window Attention
• Byte-fallback BPE tokenizer

dataset: Cabra 10k

Dataset interno para finetuning. Vamos lançar em breve.

Quantização / GGUF

Colocamos diversas versões (GGUF) quantanizadas no branch "quantanization".

Exemplo

<s> [INST] who is Elon Musk? [/INST]Elon Musk é um empreendedor, inventor e capitalista americano. Ele é o fundador, CEO e CTO da SpaceX, CEO da Neuralink e fundador do The Boring Company. Musk também é o proprietário do Twitter.</s>

Paramentros de trainamento

- learning_rate: 1e-05
- train_batch_size: 4
- eval_batch_size: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 8
- total_train_batch_size: 64
- total_eval_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.01
- num_epochs: 3

Framework

• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.14.6
• Tokenizers 0.15.2

Uso

O modelo é destinado, por agora, a fins de pesquisa. As áreas e tarefas de pesquisa possíveis incluem:

• Pesquisa sobre modelos gerativos.
• Investigação e compreensão das limitações e viéses de modelos gerativos.

Proibido para uso comercial. Somente Pesquisa.

Evals

Tasks	Version	Filter	n-shot	Metric	Value	Stderr
assin2_rte	1.1	all	15	f1_macro	0.9013	± 0.0043
		all	15	acc	0.9016	± 0.0043
assin2_sts	1.1	all	15	pearson	0.7151	± 0.0074
		all	15	mse	0.6803	± N/A
bluex	1.1	all	3	acc	0.4798	± 0.0107
		exam_id__USP_2019	3	acc	0.375	± 0.044
		exam_id__USP_2021	3	acc	0.3462	± 0.0382
		exam_id__USP_2020	3	acc	0.4107	± 0.0379
		exam_id__UNICAMP_2018	3	acc	0.4815	± 0.0392
		exam_id__UNICAMP_2020	3	acc	0.4727	± 0.0389
		exam_id__UNICAMP_2021_1	3	acc	0.413	± 0.0418
		exam_id__UNICAMP_2019	3	acc	0.42	± 0.0404
		exam_id__UNICAMP_2022	3	acc	0.5897	± 0.0456
		exam_id__USP_2022	3	acc	0.449	± 0.041
		exam_id__USP_2024	3	acc	0.6341	± 0.0434
		exam_id__UNICAMP_2024	3	acc	0.6	± 0.0422
		exam_id__USP_2023	3	acc	0.5455	± 0.0433
		exam_id__UNICAMP_2023	3	acc	0.5349	± 0.044
		exam_id__USP_2018	3	acc	0.4815	± 0.0393
		exam_id__UNICAMP_2021_2	3	acc	0.5098	± 0.0403
enem	1.1	all	3	acc	0.5843	± 0.0075
		exam_id__2010	3	acc	0.5726	± 0.0264
		exam_id__2009	3	acc	0.6	± 0.0264
		exam_id__2014	3	acc	0.633	± 0.0268
		exam_id__2022	3	acc	0.6165	± 0.0243
		exam_id__2012	3	acc	0.569	± 0.0265
		exam_id__2013	3	acc	0.5833	± 0.0274
		exam_id__2016_2	3	acc	0.5203	± 0.026
		exam_id__2011	3	acc	0.6325	± 0.0257
		exam_id__2023	3	acc	0.5778	± 0.0246
		exam_id__2016	3	acc	0.595	± 0.0258
		exam_id__2017	3	acc	0.5517	± 0.0267
		exam_id__2015	3	acc	0.563	± 0.0261
faquad_nli	1.1	all	15	f1_macro	0.6424	± 0.0138
		all	15	acc	0.6769	± 0.013
hatebr_offensive_binary	1	all	25	f1_macro	0.8361	± 0.007
		all	25	acc	0.8371	± 0.007
oab_exams	1.5	all	3	acc	0.3841	± 0.006
		exam_id__2011-03	3	acc	0.3636	± 0.0279
		exam_id__2014-14	3	acc	0.475	± 0.0323
		exam_id__2016-21	3	acc	0.4125	± 0.0318
		exam_id__2012-06a	3	acc	0.3875	± 0.0313
		exam_id__2014-13	3	acc	0.325	± 0.0303
		exam_id__2015-16	3	acc	0.425	± 0.032
		exam_id__2010-02	3	acc	0.4	± 0.0283
		exam_id__2012-08	3	acc	0.3875	± 0.0314
		exam_id__2011-05	3	acc	0.375	± 0.0312
		exam_id__2017-22	3	acc	0.4	± 0.0316
		exam_id__2018-25	3	acc	0.4125	± 0.0318
		exam_id__2012-09	3	acc	0.3636	± 0.0317
		exam_id__2017-24	3	acc	0.3375	± 0.0304
		exam_id__2016-20a	3	acc	0.3125	± 0.0299
		exam_id__2012-06	3	acc	0.425	± 0.0318
		exam_id__2013-12	3	acc	0.4375	± 0.0321
		exam_id__2016-20	3	acc	0.45	± 0.0322
		exam_id__2013-11	3	acc	0.4	± 0.0316
		exam_id__2015-17	3	acc	0.4231	± 0.0323
		exam_id__2015-18	3	acc	0.4	± 0.0316
		exam_id__2017-23	3	acc	0.35	± 0.0308
		exam_id__2010-01	3	acc	0.2471	± 0.0271
		exam_id__2011-04	3	acc	0.375	± 0.0313
		exam_id__2016-19	3	acc	0.4103	± 0.0321
		exam_id__2013-10	3	acc	0.3375	± 0.0305
		exam_id__2012-07	3	acc	0.3625	± 0.031
		exam_id__2014-15	3	acc	0.3846	± 0.0318
portuguese_hate_speech_binary	1	all	25	f1_macro	0.6187	± 0.0119
		all	25	acc	0.6322	± 0.0117