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.

• [2024-03-13] Release speechless-starcoder2-15b
• [2024-03-10] Release speechless-starcoder2-7b
• [2024-02-15] Reelase speechless-thoughts-mistral-7b-v1.0 (Change same hyperparameters)
• [2024-02-12] Release speechless-thoughts-mistral-7b
• [2024-02-10] Release speechless-sparsetral-16x7b-MoE, the MoE upgraded version of speechless-code-mistral-7b-v1.0. The MoE fine-tuning adopts Parameter-Efficient Sparsity Crafting (PESC), which is an efficient fine-tuning architecture that uses LoRA modules as expert models, similar to the concept of multi-loras.
• [2024-02-06] Release speechless-mistral-hermes-code-7b fine-tuned by speechless-thoughts-252K dataset.
• [2024-01-23] Release speechless-zephyr-code-functionary-7b
• [2024-01-15] Release speechless-nl2sql-ds-6.7b, finetune based on deepseek-coder-6.7b-base.
• [2024-01-05] Release speechless-mistral-moloras-7b, which is the static version of moloras (Mixture-of-multi-LoRAs)
• [2023/12/30] Release speechless-coder-ds-1.3b that finetune based on deepseek-coder-1.3b-base.
• [2023/12/30] Release speechless-coder-ds-6.7b that finetune based on deepseek-coder-6.7b-base.
• [2023/12/23] Uploaded the float16 version of prometheus-7b-v1.0-fp16 for instruction fine-tuning data quality assessment.
• [2023/12/23] Uploaded the float16 version of prometheus-13b-v1.0-fp16 for instruction fine-tuning data quality assessment.
• [2023/11/26] Built 7 DARE models and their respective LoRA modules.
• [2023/11/24] Start Mixture-of-Multi-LoRAs experiments. Released speechless-mistral-7b-dare-0.85
• [2023/11/21] Speechless.Tools: Released speechless-tools-7b which is trained 3 epochs on the ToolEval dataset, achieving performance comparable to ToolLlama-2-7b-v2.
• [2023/10/16] Multi-LoRAs: Load multiple LoRA modules simultaneously and automatically switch the appropriate combination of LoRA modules to generate the best answer based on user queries. uukuguy/multi-loras

The speechless-tools-7b model is fine-tuned on speechless-coding-7b-16k-tora, following the guidance of the ToolLlama project, aims to empower open-source LLMs with the ability to handle thousands of diverse real-world APIs.

speechless-tools-7b-dfs vs chatgpt-cot

speechless-tools-7b-dfs vs toolllama-dfs

Models Repositry

⭐️ My Focus 🔥🔥🔥 DL > 10k/month 🔥🔥 DL > 7K/month 🔥 DL > 3K/month

• speechless-starcoder2-15b 2024.03.13

• speechless-starcoder2-7b 2024.03.10

• speechless-thoughts-mistral-7b-v1.0 2024.02.15

  Change some hyperparameters compared to speechless-thoughts-mistral-7b.

  learning_rate=2e-4
  lora_r=64
  lora_alpha=16
  model_max_length=8192

• speechless-thoughts-mistral-7b 2024.02.12

  speechless-thoughts-mistral-7b is fine-tuned as a baseline of the speechless-sparsetral-16x7b-MoE.

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
  59.72	58.96	87.10	60.11	49.91	77.82	30.78

• ⭐ speechless-sparsetral-16x7b-MoE 2024.02.10

  The MoE upgraded version of speechless-code-mistral-7b-v1.0. The MoE fine-tuning adopts Parameter-Efficient Sparsity Crafting (PESC), which is an efficient fine-tuning architecture that uses LoRA modules as expert models, similar to the concept of multi-loras.

• ⭐ speechless-mistral-hermes-code-7b 2024.02.06

  Using the speechless-thoughts-252K dataset, it was fine-tuned based on the Mistral-7B-v0.1 base model in preparation for comparison with speechless-code-mistral-7b-v1.0.

• ⭐ speechless-zephyr-code-functionary-7b

  GGUF

• speechless-nl2sql-ds-6.7b

  GGUF

• ⭐ speechless-mistral-moloras-7b

  GPTQ AWQ GGUF by TheBloke

• ⭐ speechless-coder-ds-6.7b

  4, 5 and 8-bit GGUF models for CPU+GPU inference

• deepseek-coder-1.3b-base

• prometheus-13b-v1.0-fp16

• ⭐ prometheus-7b-v1.0-fp16

• ⭐ speechless-multi-loras-r64

• zephyr-7b-alpha-dare-0.85 2023.11.24

  A part of Mixture-of-Multi-LoRAs

• ⭐ speechless-mistral-7b-dare-0.85 2023.11.23

  GPTQ GGUF AWQ by TheBloke

  A part of Mixture-of-Multi-LoRAs

• CollectiveCognition-v1.1-Mistral-7B-dare-0.85 2023.11.23

  GPTQ GGUF AWQ by TheBloke

  A part of Mixture-of-Multi-LoRAs

• SynthIA-7B-v1.3-dare-0.85 2023.11.22

  GPTQ GGUF AWQ by TheBloke

  A part of Mixture-of-Multi-LoRAs

• airoboros-m-7b-3.1.2-dare-0.85 2023.11.22

  GPTQ GGUF AWQ by TheBloke

  A part of Mixture-of-Multi-LoRAs

• neural-chat-7b-v3-1-dare-0.85 2023.11.20

  A part of Mixture-of-Multi-LoRAs

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

  Fine-tune on the llm_agents/tora-code-7b-v1.0. The primary goal is to enhance the code generation capability of the model, thereby achieving a large-scale intelligent agent base model with good planning and reasoning abilities.

  HumanEval & MultiPL-E

  HumanEval-Python	Python	Java	JavaScript	CPP	Rust	Go	Shell	Julia	D	Lua	PHP	R
  52.44	55.96	37.84	46.93	37.48	29.01	28.99	12.11	31.47	12.05	26.52	39.25	22.09

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

  This model is a merge of 6 SOTA Mistral-7B based models. Model benchmark by sethuiyer.

• ⭐️ speechless-mistral-dolphin-orca-platypus-samantha-7b 2023.10.14

  GPTQ GGUF AWQ by TheBloke

  The subsequent version of Speechless SuperLongName based on Mistral. Ranked high in the Mistral category on the open-llm leaderboard of new evaluation indicators.

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  53.34	64.33	84.4	63.72	52.52	78.37	21.38	8.66

• 🔥🔥 speechless-tora-code-7b-v1.0 2023.10.10

  GPTQ GGUF AWQ by TheBloke

• 🔥🔥 speechless-code-mistral-orca-7b-v1.0 2023.10.10

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  55.33	59.64	82.25	61.33	48.45	77.51	8.26	49.89

• 🔥🔥 speechless-code-mistral-7b-v1.0 2023.10.10

  GPTQ GGUF AWQ by TheBloke

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  53.47	60.58	83.75	62.98	47.9	78.69	19.18	21.19

• ⭐️🔥🔥 speechless-codellama-34b-v2.0 2023.10.04

  GPTQ GGUF AWQ by TheBloke

  My current strongest code generation model supports 12 commonly used programming languages, including Python, Java, C++, Rust, Go etc. pass@1 on humaneval: 75.61, NL2SQL SQLEval: 71.43% (EM: 67.43%)

  HumanEval & MultiPL-E

  HumanEval-Python	Python	Java	JavaScript	CPP	Rust
  75.61	67.55	51.93	64.81	55.81	52.98

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  50.96	54.35	75.65	54.67	45.21	73.56	11.6	41.71

• ⭐️🔥🔥🔥 speechless-llama2-13b 2023.09.14

  GPTQ GGUF GGML (deprecated) by TheBloke

• 🔥🔥 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 by TheBloke

  Guys called it "Speechless SuperLongName". My first model with a download volume exceeding 10K last month on HuggingFace. It is said to be widely used by die-hard fans and hotly discussed on Reddit. Funny discuss on Reddit

• 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

• ⭐️🔥🔥 speechless-codellama-34b-v2.0 2023.10.04

  GPTQ GGUF AWQ by TheBloke

  My current strongest code generation model supports 12 commonly used programming languages, including Python, Java, C++, Rust, Go etc. pass@1 on humaneval: 75.61, NL2SQL SQLEval: 71.43% (EM: 67.43%)

  HumanEval & MultiPL-E

  HumanEval-Python	Python	Java	JavaScript	CPP	Rust
  75.61	67.55	51.93	64.81	55.81	52.98

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  50.96	54.35	75.65	54.67	45.21	73.56	11.6	41.71

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

• ⭐️🔥🔥🔥 speechless-codellama-platypus-13b 2023.09.13

• ⭐️🔥🔥🔥 speechless-codellama-orca-13b 2023.09.13

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

  This model is a merge of 6 SOTA Mistral-7B based models. Model benchmark by sethuiyer.

• ⭐️ speechless-mistral-dolphin-orca-platypus-samantha-7b 2023.10.14

  The subsequent version of Speechless SuperLongName based on Mistral. Ranked high in the Mistral category on the open-llm leaderboard of new evaluation indicators.

  Open LLM Language Model Evaluation Harness

  Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K	DROP
  53.34	64.33	84.4	63.72	52.52	78.37	21.38	8.66

• ⭐️🔥🔥 speechless-code-mistral-7b-v1.0 2023.10.10

  GPTQ GGUF AWQ by TheBloke

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

  Fine-tune on the llm_agents/tora-code-7b-v1.0. The primary goal is to enhance the code generation capability of the model, thereby achieving a large-scale intelligent agent base model with good planning and reasoning abilities.

  HumanEval & MultiPL-E

  HumanEval-Python	Python	Java	JavaScript	CPP	Rust	Go	Shell	Julia	D	Lua	PHP	R
  52.44	55.96	37.84	46.93	37.48	29.01	28.99	12.11	31.47	12.05	26.52	39.25	22.09

• 🔥🔥 speechless-tora-code-7b-v1.0 2023.10.10

  GPTQ GGUF AWQ by TheBloke

• ⭐️🔥🔥🔥 speechless-llama2-13b 2023.09.14

  GPTQ GGUF GGML (deprecated) by TheBloke

• ⭐️🔥 speechless-llama2-hermes-orca-platypus-wizardlm-13b 2023.09.10

  🔥🔥🔥 GPTQ GGUF AWQ by TheBloke

  Guys called it "Speechless SuperLongName". My first model with a download volume exceeding 10K last month on HuggingFace. It is said to be widely used by die-hard fans and hotly discussed on Reddit. Funny discuss on Reddit

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

python -m speechless.finetune init --task_name my_task

python -m speechless.finetune run --task_name my_task

python -m speechless.finetune merge --task_name my_task

python -m speechless.finetune backup --task_name my_task

python -m speechless.finetune list

pip install speechless

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

{
    "conversations":[
        {"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.
}

#!/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 currently supports GGUF quantification, including the following types: q4_k_m, q5_k_m, q8_0.

# quant_type: q4_km/q5_km/q8_0
python -m speechless.quant llamacpp --model_path path/to/hf/model --llamacpp_quant_type <quant_type>

Ollama is used as default backend, and litellm is used as default frontend api.

The unified classic application paradigm is to use the unified OpenAI API access interface, and the backend defaults to using the GGUF Q4_K_M quantization model.

python -m speechless.infer litellm_proxy --litellm_port 18342

python -m speechless.infer ollama_create path/to/gguf/file

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

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

LMEVAL_OUTPUT_DIR=eval_results/lm_eval/${TASK_NAME}

# lmeval
python -m speechless.eval.lmeval \
    --do_gen \
    --model hf-causal-experimental \
    --model_args pretrained=${TEST_MODEL_PATH},use_accelerate=True \
    --batch_size 4 \
    --output_path ${LMEVAL_OUTPUT_DIR} 

# lmeval_show_results
python -m speechless.eval.lmeval \
    --show_results \
    --output_path eval_results/lm_eval/${TASK_NAME} 
    --output_path ${LMEVAL_OUTPUT_DIR} 

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

HUMANEVAL_OUTPUT_DIR=eval_results/human_eval/${TASK_NAME}

# humaneval
PYTHONLIB=${SPEECHLESS_ROOT} \
python -m speechless.eval.humaneval \
    --do_gen \
    --do_eval \
    --model ${TEST_MODEL_PATH} \
    --output_dir ${HUMANEVAL_OUTPUT_DIR}

# humaneval_show_results
PYTHONLIB=${SPEECHLESS_ROOT} \
python -m speechless.eval.lmeval \
    --show_result \
    --output_path ${HUMANEVAL_OUTPU_DIR}

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

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}

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}