README.md

metadata

language:
  - ru
license: apache-2.0
datasets:
  - 0x7194633/fialka-v3-data
pipeline_tag: text-generation
model-index:
  - name: fialka-7B-v3
    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: 48.55
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          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: 71.05
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          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: 43.06
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          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: 44.79
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          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: 69.46
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          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: 1.52
            name: accuracy
        source:
          url: >-
            https://huggingface.co./spaces/HuggingFaceH4/open_llm_leaderboard?query=0x7194633/fialka-7B-v3
          name: Open LLM Leaderboard

Fialka v3.0 7B

Description

Fialka language models are trained to follow instructions and maintain communication in Russian. The third version of the model was based on the Llama 2 model, which was specifically trained on a large corpus of Russian, allowing Fialka to generate its responses more accurately.

Usage

The model has a query format as in zephyr.

<|user|>
Что такое мем?</s>
<|assistant|>
Мем - это культурный феномен, который может быть описан как информационный вирус. Он представляет собой абстрактный элемент культуры или идеи в виде изображения (обычно с изображением человека) и его краткого текста на основе которого можно создать мемы различных форматов: от графических до текстовых шуток; мема могут использоваться для передачи определенных идей через интернет-сообщество без необходимости использовать слова напрямую!</s>

Check out the space to use the model in UI without downloading.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	46.40
AI2 Reasoning Challenge (25-Shot)	48.55
HellaSwag (10-Shot)	71.05
MMLU (5-Shot)	43.06
TruthfulQA (0-shot)	44.79
Winogrande (5-shot)	69.46
GSM8k (5-shot)	1.52