language:
- bg
- en
license: apache-2.0
library_name: transformers
tags:
- mistral
- instruct
- bggpt
- insait
base_model: mistralai/Mistral-7B-v0.1
pipeline_tag: text-generation
model-index:
- name: BgGPT-7B-Instruct-v0.2
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: AI2 Reasoning Challenge (25-Shot)
type: ai2_arc
config: ARC-Challenge
split: test
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 60.58
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (10-Shot)
type: hellaswag
split: validation
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 82.18
name: normalized accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU (5-Shot)
type: cais/mmlu
config: all
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 60.5
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (0-shot)
type: truthful_qa
config: multiple_choice
split: validation
args:
num_few_shot: 0
metrics:
- type: mc2
value: 54.63
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Winogrande (5-shot)
type: winogrande
config: winogrande_xl
split: validation
args:
num_few_shot: 5
metrics:
- type: acc
value: 76.48
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: GSM8k (5-shot)
type: gsm8k
config: main
split: test
args:
num_few_shot: 5
metrics:
- type: acc
value: 44.12
name: accuracy
source:
url: >-
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=INSAIT-Institute/BgGPT-7B-Instruct-v0.2
name: Open LLM Leaderboard
INSAIT-Institute/BgGPT-7B-Instruct-v0.2
Meet BgGPT-7B, a Bulgarian language model trained from mistralai/Mistral-7B-v0.1. BgGPT is distributed under Apache 2.0 license.
This model was created by INSAIT Institute, part of Sofia University, in Sofia, Bulgaria.
This is an improved version of the model - v0.2.
Model description
The model is continously pretrained to gain its Bulgarian language and culture capabilities using multiple datasets, including Bulgarian web crawl data, a range of specialized Bulgarian datasets sourced by INSAIT Institute, and machine translations of popular English datasets. This Bulgarian data was augmented with English datasets to retain English and logical reasoning skills.
The model's tokenizer has been extended to allow for a more efficient encoding of Bulgarian words written in Cyrillic. This not only increases throughput of Cyrillic text but also performance.
Instruction format
In order to leverage instruction fine-tuning, your prompt should be surrounded by [INST] and [/INST] tokens.
The very first instruction should begin with a begin of sequence token <s>. Following instructions should not.
The assistant generation will be ended by the end-of-sequence token.
E.g.
text = "<s>[INST] Кога е основан Софийският университет? [/INST]"
"Софийският университет „Св. Климент Охридски“ е създаден на 1 октомври 1888 г.</s> "
"[INST] Кой го е основал? [/INST]"
This format is available as a chat template via the apply_chat_template() method:
Benchmarks
The model comes with a set of Benchmarks that are translations of the corresponding English-benchmarks. These are provided at https://github.com/insait-institute/lm-evaluation-harness-bg
As this is an improved version over version 0.1 of the same model and we include benchmark comparisons.
Summary
- Finetuned from: mistralai/Mistral-7B-v0.1
- Model type: Causal decoder-only transformer language model
- Language: Bulgarian and English
- License: Apache 2.0
- Contact: bggpt@insait.ai
Use in 🤗Transformers
First install direct dependencies:
pip install transformers torch accelerate
If you want faster inference using flash-attention2, you need to install these dependencies:
pip install packaging ninja
pip install flash-attn
Then load the model in transformers:
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
model="INSAIT-Institute/BgGPT-7B-Instruct-v0.2",
device_map="auto",
torch_dtype=torch.bfloat16,
use_flash_attn_2=True # optional
)
Use with GGML / llama.cpp
The model in GGUF format INSAIT-Institute/BgGPT-7B-Instruct-v0.2-GGUF
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 63.08 |
| AI2 Reasoning Challenge (25-Shot) | 60.58 |
| HellaSwag (10-Shot) | 82.18 |
| MMLU (5-Shot) | 60.50 |
| TruthfulQA (0-shot) | 54.63 |
| Winogrande (5-shot) | 76.48 |
| GSM8k (5-shot) | 44.12 |