Merge pull request #83 from fractalego/development

Development

.github/workflows/development-tests1.yml

@@ -27,7 +27,7 @@ jobs:
     - name: Install dependencies
       run: |
         sudo apt-get update
-        sudo apt-get install portaudio19-dev
+        sudo apt-get install portaudio19-dev python3-packaging
         python -m pip install --upgrade pip
         python -m pip install flake8 pytest
         if [ -f requirements.txt ]; then pip install -r requirements.txt; fi

README.md

@@ -1,4 +1,4 @@
-# WAFL 0.0.70 [![Tests](https://github.com/fractalego/wafl/actions/workflows/development-tests1.yml/badge.svg)](https://github.com/fractalego/wafl/actions/workflows/development-tests1.yml)[![Docs](https://readthedocs.org/projects/wafl/badge/?version=latest)](https://wafl.readthedocs.io/en/latest/)
+# WAFL 0.0.80 [![Tests](https://github.com/fractalego/wafl/actions/workflows/development-tests1.yml/badge.svg)](https://github.com/fractalego/wafl/actions/workflows/development-tests1.yml)[![Docs](https://readthedocs.org/projects/wafl/badge/?version=latest)](https://wafl.readthedocs.io/en/latest/)
 
 Introduction
 ============
@@ -46,16 +46,11 @@ Please see the examples in the following chapters.
 
 
 ## LLM side (needs a GPU)
-
-The second part is a machine that runs on a machine accessible from the interface side.
-The initial configuration is for a local deployment of language models. 
-No action is needed to run WAFL if you want to run it as a local instance.
-
-However, a multi-user setup will benefit for a dedicated server. 
-In this case, a docker image can be used
+The second part (LLM side) is a model server for the speech-to-text model, the LLM, the embedding system, and the text-to-speech model.
+A docker image can be used to run it as in the following:
 
 ```bash
-$ docker run -p8080:8080 --env NVIDIA_DISABLE_REQUIRE=1 --gpus all fractalego/wafl-llm:latest
+$ docker run -p8080:8080 --env NVIDIA_DISABLE_REQUIRE=1 --gpus all fractalego/wafl-llm:0.80
 ```
 
 The interface side has a `config.json` file that needs to be filled with the IP address of the LLM side.

datasets/create_rules_dataset.py

@@ -0,0 +1,45 @@
+import asyncio
+import pandas as pd
+
+from wafl.config import Configuration
+from wafl.connectors.remote.remote_llm_connector import RemoteLLMConnector
+
+
+def get_prompt(df, theme):
+    prompt = ""
+    for _, row in df.sample(9).iterrows():
+        prompt += (
+            f"""
+<task>
+Create a plausible dialogue about the theme \"{row["Theme"]}\" based on the following summary and rules.
+
+The rules are as follows:
+{row["Rules"]}
+
+The conversation goes as follows:
+{row["Conversation"]}
+</task>
+        """.strip()
+            + "\n\n"
+        )
+
+    return (
+        prompt
+        + f'<task>\nCreate plausible dialogue about the theme "{theme}" based on the following summary and rules.\n\nThe rules are as follows:\n'
+    )
+
+
+if __name__ == "__main__":
+    config = Configuration.load_local_config()
+    remote_llm_connector = RemoteLLMConnector(
+        config.get_value("llm_model"), last_strings=["</task>"]
+    )
+
+    df = pd.read_csv("data/complex_instructions.csv")
+    theme = "playing a song that the user likes"
+    prompt = get_prompt(df, theme)
+    print(
+        asyncio.run(
+            remote_llm_connector.predict(prompt, temperature=0.5, num_tokens=1500)
+        )
+    )

datasets/train_llm_on_rules_dataset.py

@@ -0,0 +1,122 @@
+import random
+
+import pandas as pd
+from datasets import Dataset
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    TrainingArguments,
+    Trainer,
+    DataCollatorForLanguageModeling,
+)
+
+model_name_or_path = "mistralai/Mistral-7B-Instruct-v0.1"
+max_length = 1024 + 512
+
+
+def get_prompts(df):
+    prompts = []
+    for _, row in df.sample(frac=1).iterrows():
+        memory = ""
+        if memory == "":
+            memory = "The user has no memory."
+
+        current_rule = row["Rules"]
+        rules = df.sample(random.choice([1, 2]))["Rules"].tolist() + [current_rule]
+        random.shuffle(rules)
+        rules = "\n".join(rules)
+        prompt = (
+            f"""
+The user is talking with a chatbot about the theme \"{row["Theme"]}\" based on the following summary.
+<summary>
+{memory}
+</summary>
+
+The rules are as follows:
+<rules>
+{rules}
+</rules>
+
+The conversation goes as follows:
+{row["Conversation"]}
+        """.strip()
+            + "\n\n"
+        )
+        prompts.append(prompt)
+
+    return prompts
+
+
+def preprocess_function(sample):
+    model_inputs = tokenizer(
+        sample["prompt"],
+        return_tensors="pt",
+        max_length=max_length,
+        padding="max_length",
+    )
+    labels = tokenizer(
+        sample["prompt"],
+        return_tensors="pt",
+        max_length=max_length,
+        padding="max_length",
+    )
+
+    model_inputs["labels"] = labels["input_ids"]
+    return model_inputs
+
+
+def model_init():
+    model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
+    parameters = model.parameters()
+    for parameter in parameters:
+        parameter.requires_grad = False
+
+    model.model.enable_input_require_grads()
+    model.lm_head.training = True
+    for index in range(len(model.model.layers)):
+        model.model.layers[index].self_attn.k_proj.training = True
+
+    return model
+
+
+def create_dataset_from_file(filepath):
+    df = pd.read_csv(filepath)
+    prompts = get_prompts(df)
+    return Dataset.from_dict({"prompt": prompts})
+
+
+if __name__ == "__main__":
+    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+    tokenizer.pad_token = tokenizer.eos_token
+    dataset = create_dataset_from_file("data/complex_instructions.csv")
+    train_dataset = dataset.map(
+        preprocess_function, batched=True, batch_size=1, num_proc=4
+    )
+    data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
+    learning_rate = 1e-6
+    output_dir_name = f"checkpoint_lr{learning_rate}"
+    training_args = TrainingArguments(
+        output_dir=output_dir_name,
+        per_device_train_batch_size=1,
+        per_device_eval_batch_size=1,
+        evaluation_strategy="steps",
+        use_cpu=True,
+        learning_rate=learning_rate,
+        num_train_epochs=2,
+        logging_steps=200,
+        eval_steps=200,
+        save_total_limit=1,
+    )
+    model = model_init()
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        train_dataset=train_dataset,
+    )
+    trainer.train()
+    trainer.save_model("wafl-mistral")
+    model = trainer.model
+    model.push_to_hub("fractalego/wafl-mistral")
+    tokenizer.push_to_hub("fractalego/wafl-mistral")

documentation/build/html/_sources/examples.rst.txt

@@ -4,5 +4,7 @@ Examples
 .. toctree::
    :maxdepth: 2
 
-   wafl_init
-   directory_structure
+   simple_rule
+   rule_with_examples
+   rules_with_execute_command
+   rules_with_remember_command

documentation/build/html/_sources/index.rst.txt

@@ -3,7 +3,7 @@
    You can adapt this file completely to your liking, but it should at least
    contain the root `toctree` directive.
 
-Welcome to WAFL's 0.0.45 documentation!
+Welcome to WAFL's 0.0.80 documentation!
 =======================================
 
 .. toctree::
@@ -12,9 +12,10 @@ Welcome to WAFL's 0.0.45 documentation!
 
    introduction
    installation
+   initialization
+   configuration
    running_WAFL
-   query_processing_pipeline
-   rules
+   facts_and_rules
    examples
    license
 

documentation/build/html/_sources/introduction.rst.txt

@@ -5,8 +5,7 @@ WAFL is a framework for personal agents.
 It integrates Large language models, speech recognition and text to speech.
 
 This framework combines Large Language Models and rules to create a predictable behavior.
-Specifically, instead of organising the work of an LLM into a chain of thoughts,
-WAFL intends to organise its behavior into inference trees.
+A set of rules is used to define the behavior of the agent, supporting function calling and a working memory.
 
 WAFL is a work in progress.
 The current version requires the user to specify the rules to follow.

documentation/build/html/_sources/license.rst.txt

@@ -3,7 +3,7 @@ License
 
 This software is licensed under the MIT License:
 
-Copyright (c) 2023 [email protected]
+Copyright (c) 2024 [email protected]
 
 Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: