mini-mistral-openhermes-2.5-chatml-test

Model Card for neoncortex/mini-mistral-openhermes-2.5-chatml-test

A tiny Mistral model trained as an experiment on teknium/OpenHermes-2.5.

Model Details

A 63M parameter auto-regressive LM using Mistral architecture as a base.

• Multi-query Attention instead of Grouped-query Attention.
• Sliding window is disabled.
• Modified ChatML instead of Mistral chat template - TL;DR I used '<|im_start|>human' instead of '<|im_start|>user'

Model Description

Just doing it to see what happens.

It'll take about 40 to 45 hours to train on two Nvidia RTX 3060 12GB.

It uses ChatML for the chat template, but I messed up the template in the dataset, using '<|im_start|>human' instead of '<|im_start|>user'. ¯_(ツ)_/¯ So, here's the bits:

{%- set ns = namespace(found=false) -%}
{%- for message in messages -%}
    {%- if message['role'] == 'system' -%}
        {%- set ns.found = true -%}
    {%- endif -%}
{%- endfor -%}
{%- for message in messages %}
    {%- if message['role'] == 'system' -%}
        {{- '<|im_start|>system\n' + message['content'].rstrip() + '<|im_end|>\n' -}}
    {%- else -%}
        {%- if message['role'] == 'human' -%}
            {{-'<|im_start|>human\n' + message['content'].rstrip() + '<|im_end|>\n'-}}
        {%- else -%}
            {{-'<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' -}}
        {%- endif -%}
    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{-'<|im_start|>assistant\n'-}}
{%- endif -%}

• Developed by: RoboApocalypse
• Funded by: RoboApocalypse
• Shared by: RoboApocalypse
• Model type: Mistral
• Language(s) (NLP): English, maybe others idk
• License: OpenRAIL

Model Sources

Exclusively available right here on HuggingFace!

• Repository: https://huggingface.co/neoncortex/mini-mistral-openhermes-2.5-chatml-test

Uses

None

Out-of-Scope Use

This model won't work well for pretty much everything, probably.

Preprocessing

Format the OpenHermes 2.5 dataset with ChatML.

Training Hyperparameters

• Training regime: bf16 mixed precision

Evaluation

I tried to run evals but the eval suite just laughed at me.

Model Examination

Don't be rude.

Environmental Impact

• Hardware Type: 2 x NVIDIA RTX 3060 12GB
• Hours used: ~45 x 2.
• Carbon Emitted: [TBA]

Compute Infrastructure

I trained it on my PC with no side on it because I like to watch the GPUs do their work.

Hardware

2 x Nvidia RTX 3060 12GB

Software

The wonderful free stuff at HuggingFace (https://huggingface.co)[https://huggingface.co]: transformers, datasets, trl