speechless 0.6.0

LLM based agents with proactive interactions, long-term memory, external tool integration, and local deployment capabilities.

Speechless LLM based Agents

LLM based agents with proactive interactions, long-term memory, external tool integration, and local deployment capabilities.

Speechless.AI is committed to integrating the superior language processing and deep reasoning capabilities of large language models into practical business applications. By enhancing the model's language understanding, knowledge accumulation, and text creation abilities, and introducing long-term memory, external tool integration, and local deployment, our aim is to establish an intelligent collaborative partner that can independently interact, continuously evolve, and closely align with various business scenarios.

• Firstly, we focus on building a large model with enhanced reasoning capabilities, ensuring its outstanding performance in language processing and logical analysis.

• Next, we design and implement an efficient operational framework for the intelligent entity. This framework not only supports rapid deployment and invocation of the model but also boasts features like autonomous interaction, real-time feedback adjustment, context awareness, and long-term memory. For instance, in customer service scenarios, the intelligent entity can provide more precise and personalized responses based on a user's historical interactions and current context. In content recommendation scenarios, it can dynamically adjust its strategies by capturing real-time shifts in user interests.

• Ultimately, we integrate it with real business scenarios, ensuring that the intelligent entity seamlessly aligns with various business processes, delivering tangible value to enterprises.

Models

Models Repositry

• speechless-coding-7b-16k-tora 2023.11.01

• speechless-mistral-six-in-one-7b 2023.10.15

• speechless-mistral-dolphin-orca-platypus-samantha-7b 2023.10.14 GPTQ GGUF AWQ

• speechless-tora-code-7b-v1.0 2023.10.10 GPTQ GGUF AWQ

• speechless-code-mistral-7b-v1.0 2023.10.10 GPTQ GGUF AWQ

• speechless-codellama-34b-v2.0 2023.10.04 GPTQ GGUF AWQ

• speechless-llama2-13b 2023.09.14 GPTQ GGUF GGML (deprecated)

• speechless-codellama-airoboros-orca-platypus-13b 2023.09.19

• speechless-codellama-dolphin-orca-platypus-34b 2023.09.14

• speechless-llama2-dolphin-orca-platypus-13b 2023.09.16

• speechless-codellama-34b-v1.0 2023.09.14

• speechless-codellama-platypus-13b 2023.09.13

• speechless-codellama-orca-13b 2023.09.13

• speechless-llama2-hermes-orca-platypus-wizardlm-13b 2023.09.10 GPTQ GGUF AWQ

• speechless-llama2-hermes-orca-platypus-13b 2023.09.02

• speechless-llama2-luban-orca-platypus-13b 2023.09.01

• speechless-hermes-coig-lite-13b 2023.08.22

CodeLlama based Models

• speechless-codellama-34b-v2.0 2023.10.04 GPTQ GGUF AWQ

• speechless-codellama-airoboros-orca-platypus-13b 2023.09.19

Mistral based Models

• speechless-mistral-six-in-one-7b 2023.10.15
• speechless-mistral-dolphin-orca-platypus-samantha-7b 2023.10.14 GPTQ GGUF AWQ
• speechless-code-mistral-7b-v1.0 2023.10.10 GPTQ GGUF AWQ

Tora based Models

• speechless-coding-7b-16k-tora 2023.11.01
• speechless-tora-code-7b-v1.0 2023.10.10 GPTQ GGUF AWQ

Llama2 based Models

• speechless-llama2-hermes-orca-platypus-wizardlm-13b 2023.09.10 GPTQ GGUF AWQ
• speechless-mistral-dolphin-orca-platypus-samantha-7b 2023.10.14 GPTQ GGUF AWQ

Datasets

• jondurbin/airoboros-2.2.1
• Open-Orca/OpenOrca
• garage-bAInd/Open-Platypus
• WizardLM/WizardLM_evol_instruct_V2_196k
• ehartford/dolphin
• ehartford/samantha-data

speechless.finetune

Install speechless

pip install speechless

Prepare train dataset

The training dataset is a jsonl file, with each line containing a JSON formatted instruction data. The data format is as follows:

{
    "dialog":[
        {"from": "human", "value": "Human's Instruction"},
        {"from": "assistant", "value": "Assistant's response"}
    ],
    "prompt_type": "alpaca", # Current support 'alpaca', 'toolllama-multi-rounds', default is 'alpaca' if prompt_type set to empty.
    "system_prompt": "", # Use alpaca system prompt if system_prompt filed is empty, otherwise use it as system prompt of this instruction.
    "category": "my_category", # User customized category, can be anythings.
}

Run Fine-tune

#!/bin/bash
SCRIPT_PATH=$(cd $(dirname ${BASH_SOURCE[0]}); pwd)

# -------------------- Model --------------------
export MODELS_ROOT_DIR=/opt/local/llm_models/huggingface.co
export BASE_MODEL_PATH=${MODELS_ROOT_DIR}/llm_agents/tora-code-7b-v1.0
export TEST_MODEL_PATH=${MODELS_ROOT_DIR}/speechlessai/$(basename ${PWD})

# -------------------- Dataset --------------------
export SPEECHLESS_DATA_DIR=/opt/local/datasets/speechless_data
export DATASET=${SPEECHLESS_DATA_DIR}/speechless-toolbench-multi-rounds.jsonl
export DATASET_FORMAT=dialog

# -------------------- Environment --------------------
export OUTPUT_DIR=./outputs
export RAY_memory_monitor_refresh_ms=0

# -------------------- Task --------------------
export TASK_NAME=$(basename ${TEST_MODEL_PATH})
export TASK_CHECKPOINT_DIR=${OUTPUT_DIR}
export WANDB_PROJECT=${TASK_NAME}

# -------------------- Train --------------------
export SAVE_STEPS=10
export EVAL_STEPS=10
export WARMUP_STEPS=10
export MAX_EVAL_SAMPLES=200
export EVAL_DATASET_SIZE=0.005
export GROUP_BY_LENGTH=False
export LR_SCHEDULER_TYPE=cosine
export LEARNING_RATE=2e-4

export BITS=4
export LORA_R=32
export LORA_ALPHA=256

export MODEL_MAX_LENGTH=32768
export ROPE_THETA=1000000
export SLIDING_WINDOW=8192

export NUM_GPUS=2
export NUM_TRAIN_EPOCHS=3

export SAVE_STRATEGY=epoch
export SAVE_TOTAL_LIMIT="--save_total_limit ${NUM_TRAIN_EPOCHS}"

export PER_DEVICE_TRAIN_BATCH_SIZE=2
export GRADIENT_ACCUMULATION_STEPS=16
export MAX_MEMORY_MB=32000

PYTHONPATH=${SPEECHLESS_ROOT} \
torchrun --nnodes=1 --nproc_per_node=${NUM_GPUS} \
    -m speechless.finetune.finetune_dialog \
    --task_name ${TASK_NAME} \
    --run_name $(date +%Y%m%d-%H%M%S) \
    --model_name_or_path ${BASE_MODEL_PATH} \
    --output_dir ${OUTPUT_DIR} \
    --num_train_epochs ${NUM_TRAIN_EPOCHS} \
    --data_seed 10042 \
    --save_strategy ${SAVE_STRATEGY} \
    ${SAVE_TOTAL_LIMIT} \
    --evaluation_strategy steps \
    --eval_dataset_size ${EVAL_DATASET_SIZE} \
    --save_steps ${SAVE_STEPS} \
    --eval_steps ${EVAL_STEPS} \
    --warmup_steps ${WARMUP_STEPS} \
    --max_train_samples ${MAX_TRAIN_SAMPLES} \
    --max_eval_samples ${MAX_EVAL_SAMPLES} \
    --dataloader_num_workers 3 \
    --logging_strategy steps \
    --logging_steps 1 \
    --report_to tensorboard \
    --remove_unused_columns False \
    --do_train \
    --max_memory_MB ${MAX_MEMORY_MB} \
    --bits ${BITS} \
    --lora_r ${LORA_R} \
    --lora_alpha ${LORA_ALPHA} \
    --lora_dropout 0.05 \
    --lora_modules all \
    --double_quant \
    --quant_type nf4 \
    --bf16 \
    --sliding_window ${SLIDING_WINDOW} \
    --rope_theta ${ROPE_THETA} \
    --dataset ${DATASET} \
    --dataset_format ${DATASET_FORMAT} \
    --max_new_tokens ${MODEL_MAX_LENGTH} \
    --model_max_len ${MODEL_MAX_LENGTH} \
    --per_device_train_batch_size ${PER_DEVICE_TRAIN_BATCH_SIZE} \
    --gradient_accumulation_steps ${GRADIENT_ACCUMULATION_STEPS} \
    --per_device_eval_batch_size 1 \
    --learning_rate ${LEARNING_RATE} \
    --lr_scheduler_type ${LR_SCHEDULER_TYPE} \
    --weight_decay 0.0 \
    --seed 10042 \
    --optim paged_adamw_8bit \
    --gradient_checkpointing True \
    --group_by_length ${GROUP_BY_LENGTH} \
    --ddp_find_unused_parameters False \
    --force_remove_overlength_samples False \
    --flash_attention True 

speechless.api.server

python -m speechless.api.server \
    start \
    --model ${TASK_MODEL_PATH} \
    --backbone vllm \
    --host 0.0.0.0 \
    --port 5001

speechless.eval

Speechless supports HumanEval, MultiPL-E, SQLEval, lm-evaluation-harness.

lm-evluation-harness

python -m speechless.eval.lmeval \
   genrate \
    --model ${TASK_MODEL_PATH} \
    --output_dir ${EVAL_OUTPUT_DIR} \

python -m speechless.eval.lmeval \
    eval \
    -eval_dir ${EVAL_OUTPUT_DIR}

bigcode-evaluation-harness

docker pull ghcr.io/bigcode-project/evaluation-harness
docker tag ghcr.io/bigcode-project/evaluation-harness evaluation-harness

HumanEval

Execute the HumanEval geenrate command on the GPU server where the model is located.

python -m speechless.eval.humaneval \
    genrate \
    --model ${TASK_MODEL_PATH} \
    --output_dir ${EVAL_OUTPUT_DIR} \

python -m speechless.eval.humaneval \
    eval \
    --eval_dir ${EVAL_OUTPUT_DIR}

MultiPL-E

docker pull ghcr.io/bigcode-project/evaluation-harness-multiple
docker tag ghcr.io/bigcode-project/evaluation-harness-multiple evaluation-harness-multiple

python -m speechless.eval.multiple \
    genrate \
    --name ${TASK_MODEL_PATH} \
    --output_dir_prefix ${EVAL_OUTPUT_DIR} \

python -m speechless.eval.multiple \
    eval \
    --results_dir ${EVAL_OUTPUT_DIR}

SQLEval

python -m speechless.eval.sqleval \
    genrate \
    --model ${TASK_MODEL_PATH} \
    --output_dir ${EVAL_OUTPUT_DIR} \

python -m speechless.eval.sqleval \
    eval \
    --eval_dir ${EVAL_OUTPUT_DIR}