---
license: gpl-3.0
model-index:
- name: 34b-beta
  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: 70.56
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      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: 84.2
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      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: 85.6
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      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: 58.38
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      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: 81.29
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      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: 58.23
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=CausalLM/34b-beta
      name: Open LLM Leaderboard
---
# CausalLM 34B β

## PROMPT FORMAT:
[chatml](https://github.com/openai/openai-python/blob/main/chatml.md)

There are some issues with the model weights in terms of precision. In the next version update, we will roll back some progress and retrain to fix these issues as soon as possible.

**Please note:** Do not use "accelerated inference frameworks" like **VLLM** temporarily. Instead, use Transformers for inference. Otherwise, due to precision issues, the output quality will be significantly degraded. If you need faster inference, you can consider using the q8_0 quantization (faster and better than bf16 vllm for this model only) with llama.cpp temporarily or wait for the official version.
To be fixed in the upcoming next version update.

**no repetition_penalty!**

Please do not use wikitext for quantization calibration because all wikitext have been re-aligned on synthetic dataset, and its distribution differs significantly from the original wikitext.

## MT-Bench: 8.5

![mt-bench](https://cdn-uploads.huggingface.co/production/uploads/63468a143ea42ee2cb49ddd1/2vv2_nGbfWuOM8jwy40dn.png)

## Some contamination detection if you want to check:

| Models                    | MMLU (ref: llama7b) | TBA  |
| ------------------------- | ------------------- | ---- |
| microsoft/Orca-2-7b       | 0.77                |      |
| mistralai/Mistral-7B-v0.1 | 0.46                |      |
| **CausalLM/34b-beta**     | **0.38**            |      |
| 01-ai/Yi-6B-200K          | 0.3                 |      |

data from https://huggingface.co/spaces/Yeyito/llm_contamination_detector

It should be *safe*. It was not trained on the benchmark, but the contamination of the training dataset is unavoidable due to cost constraints.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_CausalLM__34b-beta)

|             Metric              |Value|
|---------------------------------|----:|
|Avg.                             |73.04|
|AI2 Reasoning Challenge (25-Shot)|70.56|
|HellaSwag (10-Shot)              |84.20|
|MMLU (5-Shot)                    |85.60|
|TruthfulQA (0-shot)              |58.38|
|Winogrande (5-shot)              |81.29|
|GSM8k (5-shot)                   |58.23|