Dracones/c4ai-command-r-v01_exl2_3.0bpw-rpcal

c4ai-command-r-v01 - EXL2 3.0bpw

This is a 3.0bpw EXL2 quant of CohereForAI/c4ai-command-r-v01

Details about the model can be found at the above model page.

EXL2 Version

These quants were made with exllamav2 version 0.0.18. Quants made on this version of EXL2 may not work on older versions of the exllamav2 library.

If you have problems loading these models, please update Text Generation WebUI to the latest version.

RP Calibrated

The rpcal quants were made using data/PIPPA-cleaned/pippa_raw_fix.parquet for calibration.

Perplexity Scoring

Below are the perplexity scores for the EXL2 models. A lower score is better.

Stock Quants

Quant Level	Perplexity Score
8.0	6.4436
7.0	6.4372
6.0	6.4391
5.0	6.4526
4.5	6.4629
4.0	6.5081
3.5	6.6301
3.0	6.7974

RP Calibrated Quants

Quant Level	Perplexity Score
8.0	6.4331
7.0	6.4347
6.0	6.4356
5.0	6.4740
4.5	6.4875
4.0	6.5039
3.5	6.6928
3.0	6.8913

EQ Bench

Here are the EQ Bench scores for the EXL2 quants using Alpaca, ChatML, Command-R and Command-R-Plus prompt templates. A higher score is better.

Quants

Quant Size	ChatML	Alpaca	Command-R	Command-R-Plus
8.0	47.28	56.67	58.46	58.49
7.0	46.86	57.5	57.29	57.91
6.0	48.61	56.5	57.8	58.64
5.0	48.48	54.64	57.14	56.63
4.5	48.1	57.75	57.08	56.7
4.0	50.99	53.41	57.46	57.99
3.5	52.72	56.68	60.91	60.91
3.0	39.19	36.45	49.17	49.68

RP Calibrated Quants

Quant Size	ChatML	Alpaca	Command-R	Command-R-Plus
8.0	48.42	56.23	58.41	58.41
7.0	48.47	57.01	57.85	57.67
6.0	50.93	58.33	60.32	59.83
5.0	50.29	55.28	58.96	59.23
4.5	46.63	55.01	57.7	59.24
4.0	47.13	49.76	54.76	55.5
3.5	52.98	56.39	59.19	58.32
3.0	47.94	50.36	54.89	53.61

Command-R-Plus Template

This is the Command-R-Plus template yaml that was used in EQ bench(which uses Text Generation Web UI yaml templates). It adds BOS_TOKEN into the starter prompt.

text-generation-webui/instruction-templates/Command-R-Plus.yaml:

instruction_template: |-
  {%- if messages[0]['role'] == 'system' -%}
      {%- set loop_messages = messages[1:] -%}
      {%- set system_message = messages[0]['content'] -%}
  {%- elif false == true -%}
      {%- set loop_messages = messages -%}
      {%- set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' -%}
  {%- else -%}
      {%- set loop_messages = messages -%}
      {%- set system_message = false -%}
  {%- endif -%}
  {%- if system_message != false -%}
      {{ '<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}
  {%- endif -%}
  {%- for message in loop_messages -%}
      {%- set content = message['content'] -%}
      {%- if message['role'] == 'user' -%}
          {{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}
      {%- elif message['role'] == 'assistant' -%}
          {{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}
      {%- endif -%}
  {%- endfor -%}
  {%- if add_generation_prompt -%}
      {{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}
  {%- endif -%}

Perplexity Script

This was the script used for perplexity testing.

#!/bin/bash

# Activate the conda environment
source ~/miniconda3/etc/profile.d/conda.sh
conda activate exllamav2

# Set the model name and bit size
MODEL_NAME="c4ai-command-r-v01"
BIT_PRECISIONS=(8.0 7.0 6.0 5.0 4.5 4.0 3.5 3.0)

# Print the markdown table header
echo "| Quant Level | Perplexity Score |"
echo "|-------------|------------------|"

for BIT_PRECISION in "${BIT_PRECISIONS[@]}"
do
  MODEL_DIR="models/${MODEL_NAME}_exl2_${BIT_PRECISION}bpw"
  # MODEL_DIR="models/${MODEL_NAME}_exl2_${BIT_PRECISION}bpw-rpcal"
  if [ -d "$MODEL_DIR" ]; then
    output=$(python test_inference.py -m "$MODEL_DIR" -gs 22,24 -ed data/wikitext/wikitext-2-v1.parquet)
    score=$(echo "$output" | grep -oP 'Evaluation perplexity: \K[\d.]+')
    echo "| $BIT_PRECISION | $score |"
  fi
done

Quant Details

This is the script used for quantization.

#!/bin/bash

# Activate the conda environment
source ~/miniconda3/etc/profile.d/conda.sh
conda activate exllamav2

# Set the model name and bit size
MODEL_NAME="c4ai-command-r-v01"

# Define variables
MODEL_DIR="models/$MODEL_NAME"
OUTPUT_DIR="exl2_$MODEL_NAME"
MEASUREMENT_FILE="measurements/$MODEL_NAME.json"
# CALIBRATION_DATASET="data/PIPPA-cleaned/pippa_raw_fix.parquet"

# Create the measurement file if needed
if [ ! -f "$MEASUREMENT_FILE" ]; then
    echo "Creating $MEASUREMENT_FILE"
    # Create directories
    if [ -d "$OUTPUT_DIR" ]; then
        rm -r "$OUTPUT_DIR"
    fi
    mkdir "$OUTPUT_DIR"

    # python convert.py -i $MODEL_DIR -o $OUTPUT_DIR -nr -om $MEASUREMENT_FILE -c $CALIBRATION_DATASET
    python convert.py -i $MODEL_DIR -o $OUTPUT_DIR -nr -om $MEASUREMENT_FILE
fi

# Choose one of the below. Either create a single quant for testing or a batch of them.
# BIT_PRECISIONS=(5.0)
BIT_PRECISIONS=(8.0 7.0 6.0 5.0 4.5 4.0 3.5 3.0)

for BIT_PRECISION in "${BIT_PRECISIONS[@]}"
do
    CONVERTED_FOLDER="models/${MODEL_NAME}_exl2_${BIT_PRECISION}bpw"

    # If it doesn't already exist, make the quant
    if [ ! -d "$CONVERTED_FOLDER" ]; then

        echo "Creating $CONVERTED_FOLDER"

        # Create directories
        if [ -d "$OUTPUT_DIR" ]; then
            rm -r "$OUTPUT_DIR"
        fi
        mkdir "$OUTPUT_DIR"
        mkdir "$CONVERTED_FOLDER"
        
        # Run conversion commands  
        # python convert.py -i $MODEL_DIR -o $OUTPUT_DIR -nr -m $MEASUREMENT_FILE -b $BIT_PRECISION -c $CALIBRATION_DATASET -cf $CONVERTED_FOLDER
        python convert.py -i $MODEL_DIR -o $OUTPUT_DIR -nr -m $MEASUREMENT_FILE -b $BIT_PRECISION -cf $CONVERTED_FOLDER

    fi
done