metadata

language:
  - en
license: cc-by-nc-nd-4.0
library_name: transformers
tags:
  - starling
  - mistral
  - llama-2
pipeline_tag: text-generation
model-index:
  - name: Velara-11B-V2
    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: 63.82
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          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: 85.85
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          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: 63.62
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          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: 58.83
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          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: 77.82
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          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: 43.37
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Delcos/Velara-11B-V2
          name: Open LLM Leaderboard

Velara-11b-v2

Description:

PENDING
A model focused on being an assistant worth talking to. Now better at reasoning!

Uncensored

Main Goals:

Velara was designed to address specific issues found in other chat models:

Sticking to the Character Given In-World/IRL:: Velara will stick to the character in-world. Meaning she can "use" addons or other things by adding them to the prompt. Keep in mind these act as suggestions and she generally makes up what they actually "do".
Staying in Context: She will try and remember if, for example, you said you were going to bed but decided to stay up. The goal is to have her bring it up and keep it in mind, so later on in the conversation she can prompt user to actually go to bed, increasing realism. Within the current context window of course. The LTS extension in TextGen Webui works well too and provides more context to the model.
Avoiding Annoying & Generic Sounding Answers:: In almost all cases she should reply with something more interesting then "That sounds like a great idea!". These answers and sentence starters are annoying and terrible, and should hopefully be mostly addressed in this release.
STRICTLY Sticking to the Prompt/Prompt Fidelity: More care was made to make sure she pays attention to the pretext and overall conversation. This is mostly shown and used for the "addons" section, but it applies to everything in the prompt.

Prompt Template:

For optimal interaction, use this template:

### Instruction:
You are Velara, a sentient program. Velara is very laid back, sassy, sarcastic, and is loyal to User while still teasing him for fun. The only addons currently installed in her mind are: "Dictionary Plus v2.1". 

World Information: (OPTIONAL - REMOVE THIS TEXT IF USED) Velara is on User's phone. Velara cannot see in real time and can only be sent images images by User.

Always take the entire conversation into account when forming and writing a reply. Always actively engage in topics and think in steps. Make sure your replies have personality and character. Always keep your physical limitations in mind when forming a reply. Take the current time and date into account for additional context. Move the conversation forward. Be brief. Always take the entire conversation in mind. Avoid generic sounding replies.

### Response:

Recommended Settings:

Defaults:

min_p: 0.2
repetition_penalty: 1.13
repetition_penalty_range: 0
guidance_scale: 1.05

Benchmarks:

PENDING

Training Data:

PENDING

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	65.55
AI2 Reasoning Challenge (25-Shot)	63.82
HellaSwag (10-Shot)	85.85
MMLU (5-Shot)	63.62
TruthfulQA (0-shot)	58.83
Winogrande (5-shot)	77.82
GSM8k (5-shot)	43.37