mario-rc/emotional-rlaif-ppo-meta-llama-3-8b-instruct

Emotional RLAIF PPO Meta-Llama-3-8B-Instruct

This repository contains a LoRA/PEFT adapter trained from meta-llama/Meta-Llama-3-8B-Instruct with LLaMA-Factory using Proximal Policy Optimization (PPO) for emotional response alignment.

The adapter was trained as part of an emotional RLAIF pipeline using the mario-rc/aif-emotional-generation dataset, with dialogues used for SFT and PPO, and aif_annotations preferences used for RM preference alignment.

Project repository: Mario-RC/aif-emotional-model.

Intended Use

This adapter is intended for research and experimentation with emotionally aligned dialogue generation. It should be loaded on top of the corresponding base model using PEFT.

Model Details

• Base model: meta-llama/Meta-Llama-3-8B-Instruct
• Adapter repository: mario-rc/emotional-rlaif-ppo-meta-llama-3-8b-instruct
• Adapter type: LoRA / PEFT
• Alignment method: PPO
• Training framework: LLaMA-Factory
• Prompt template: llama3
• Dataset: mario-rc/aif-emotional-generation

Released Emotional RLAIF Models

The released emotional RLAIF adapters are available on Hugging Face:

Model	Base model	Size	Alignment method	Prompt template
emotional-rlaif-dpo-gemma-2-2b-it	google/gemma-2-2b-it	2B	DPO	gemma
emotional-rlaif-ppo-gemma-2-9b-it	google/gemma-2-9b-it	9B	PPO	gemma
emotional-rlaif-dpo-gemma-2-9b-it	google/gemma-2-9b-it	9B	DPO	gemma
emotional-rlaif-ppo-glm-4-9b-chat-1m	THUDM/glm-4-9b-chat-1m	9B	PPO	glm4
emotional-rlaif-dpo-glm-4-9b-chat-1m	THUDM/glm-4-9b-chat-1m	9B	DPO	glm4
emotional-rlaif-ppo-meta-llama-3-8b-instruct	meta-llama/Meta-Llama-3-8B-Instruct	8B	PPO	llama3
emotional-rlaif-dpo-meta-llama-3-8b-instruct	meta-llama/Meta-Llama-3-8B-Instruct	8B	DPO	llama3
emotional-rlaif-ppo-llama-3.2-1b-instruct	meta-llama/Llama-3.2-1B-Instruct	1B	PPO	llama3
emotional-rlaif-dpo-llama-3.2-3b-instruct	meta-llama/Llama-3.2-3B-Instruct	3B	DPO	llama3
emotional-rlaif-ppo-mistral-7b-instruct-v0.3	mistralai/Mistral-7B-Instruct-v0.3	7B	PPO	mistral
emotional-rlaif-ppo-phi-3-small-8k-instruct	microsoft/Phi-3-small-8k-instruct	7B	PPO	phi

Training Procedure

Key hyperparameters for this adapter:

• Learning rate: 1e-5
• Epochs: 1
• Scheduler: cosine
• Warmup ratio: 0.1
• SFT data: dialogues
• PPO prompt source: dialogues
• Reward model data: aif_annotations preference pairs
• Precision: bfloat16
• Optimizer: AdamW (torch)

Evaluation

The released adapters were evaluated on the held-out emotional dialogue generation split with automatic reference-overlap metrics. These metrics measure similarity to reference responses and should be complemented with human or judge-based evaluation for emotional quality.

All released adapters:

Model	Alignment	BLEU-4	ROUGE-1	ROUGE-2	ROUGE-L
emotional-rlaif-dpo-gemma-2-2b-it	DPO	38.0977	39.3142	19.5753	33.7190
emotional-rlaif-ppo-gemma-2-9b-it	PPO	44.4333	42.5233	22.7153	37.7601
emotional-rlaif-dpo-gemma-2-9b-it	DPO	41.9904	40.8481	21.2884	35.8110
emotional-rlaif-ppo-glm-4-9b-chat-1m	PPO	45.0070	42.0043	22.8903	37.9557
emotional-rlaif-dpo-glm-4-9b-chat-1m	DPO	40.0157	39.5339	19.6798	34.3553
emotional-rlaif-ppo-meta-llama-3-8b-instruct	PPO	44.3123	43.1426	23.3597	38.3680
emotional-rlaif-dpo-meta-llama-3-8b-instruct	DPO	41.0271	41.0354	21.1472	36.0910
emotional-rlaif-ppo-llama-3.2-1b-instruct	PPO	41.2577	39.9133	20.1419	34.4066
emotional-rlaif-dpo-llama-3.2-3b-instruct	DPO	41.7391	40.7582	20.9527	35.5167
emotional-rlaif-ppo-mistral-7b-instruct-v0.3	PPO	45.9741	44.2755	25.5357	40.8223
emotional-rlaif-ppo-phi-3-small-8k-instruct	PPO	45.6238	44.3653	25.4530	40.1762

Framework Versions

• PEFT: 0.11.1
• Transformers: 4.42.x / 4.45.x training environments
• PyTorch: bfloat16 CUDA training
• LLaMA-Factory: LoRA/PEFT training workflow

Usage Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
adapter_id = "mario-rc/emotional-rlaif-ppo-meta-llama-3-8b-instruct"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

messages = [
    {"role": "user", "content": "I feel overwhelmed today. Can you respond with empathy?"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.7,
        top_p=0.9,
    )

print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

How to Use

The following example loads this adapter and runs an interactive emotional dialogue loop. Use exit to stop the chat.

import random
import sys

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

MODEL_ID = "mario-rc/emotional-rlaif-ppo-meta-llama-3-8b-instruct"


def get_turn_markers():
    return {
        'bos': '<|begin_of_text|>',
        'user_start': '<|start_header_id|>user<|end_header_id|>\n\n',
        'user_end': '<|eot_id|>',
        'assistant_start': '<|start_header_id|>assistant<|end_header_id|>\n\n',
        'assistant_end': '<|eot_id|>',
    }


def update_prompt(dialogues):
    """Build the prompt for the model based on the dialogue history."""
    markers = get_turn_markers()

    system = (
        f"{markers['bos']}You are an expert at creating dialogues.\n\n"
        "Dialogue and emotional structure:\n"
    )

    human_prompts = [d[0] for d in dialogues]
    chatbot_responses = [d[1] for d in dialogues]

    p_emo = [h[0] for h in human_prompts]
    p_utt = [h[1] for h in human_prompts]
    r1_utt = [c[1] for c in chatbot_responses]
    r2_emo = [c[2] for c in chatbot_responses]
    r2_utt = [c[3] for c in chatbot_responses]
    r3_utt = [c[5] for c in chatbot_responses]

    context = (
        "Human: (HAPPINESS) PROMPT.\n"
        "Chatbot: (HAPPINESS) RESPONSE_1. (HAPPINESS) RESPONSE_2. (NEUTRAL) RESPONSE_3.\n"
    )
    for p_e, _, _, r2_e, _, _ in zip(p_emo, p_utt, r1_utt, r2_emo, r2_utt, r3_utt):
        context += f"Human: ({p_e}) PROMPT.\n"
        context += f"Chatbot: ({p_e}) RESPONSE_1. ({r2_e}) RESPONSE_2. (NEUTRAL) RESPONSE_3.\n"
    context += "\n"

    rules = (
        "Dialogue rules:\n"
        "The response must be open-domain curated. The response should be coherent, empathetic, engaging and proactive.\n"
        "The chatbot RESPONSE is composed of 3 different sentences (RESPONSE_1, RESPONSE_2 and RESPONSE_3), separated by a period.\n"
        "Between RESPONSE_1, RESPONSE_2 and RESPONSE_3 should be a max length of 20-25 words.\n"
        "RESPONSE_3 must be open-ended to follow-up the conversation, so the Human is encouraged to answer with a full long sentence. Avoid yes/no questions.\n\n"
        "Emotional response rules:\n"
        f"RESPONSE_1 must contain a {p_emo[-1]} tone.\n"
        f"RESPONSE_2 must contain a {r2_emo[-1]} tone.\n"
        "RESPONSE_3 must contain a NEUTRAL tone.\n\n"
        "Answer in a single turn to Human. Follow exactly the emotional structure and the emotional and dialogue rules."
        f"{markers['user_start']}"
    )

    completion = ""
    for idx, (p_e, p_u, r1_u, r2_e, r2_u, r3_u) in enumerate(zip(p_emo, p_utt, r1_utt, r2_emo, r2_utt, r3_utt)):
        completion += f"({p_e}) {p_u}{markers['user_end']}{markers['assistant_start']}"
        if idx != len(p_emo) - 1:
            completion += f"({p_e}) {r1_u} ({r2_e}) {r2_u} (NEUTRAL) {r3_u}{markers['assistant_end']}{markers['user_start']}"

    return system + context + rules + completion


class Chatbot:
    def __init__(self, dialogue_language="es"):
        self.dialogue_language = dialogue_language
        self.device = "cuda" if torch.cuda.is_available() else "cpu"

        self.tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        self.model = AutoPeftModelForCausalLM.from_pretrained(
            MODEL_ID,
            torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
            device_map="auto" if torch.cuda.is_available() else None,
        )
        if not torch.cuda.is_available():
            self.model = self.model.to(self.device)
        self.model.eval()

    def split_emo_chatbot(sentence):
        """Extract emotions and utterances from a chatbot response."""
        response_pos_ini = [i for i, c in enumerate(sentence) if c == "("]
        response_pos_end = [i for i, c in enumerate(sentence) if c == ")"]
        response_r1_utt = sentence[response_pos_end[0] + 2:response_pos_ini[1]].strip()
        response_r2_utt = sentence[response_pos_end[1] + 2:response_pos_ini[2]].strip()
        response_r3_utt = sentence[response_pos_end[2] + 2:].lstrip()
        return response_r1_utt, response_r2_utt, response_r3_utt

    def select_dialogue(self, dialogue_language):
        """Return a list of example dialogues for the given language."""
        if dialogue_language == "en":
            dialogue_base = [
                [["HAPPINESS", "Hi, who are you?"],
                 ["HAPPINESS", "Hi! I'm Ray, a social personal assistant robot with emotions.", "HAPPINESS", "I'm here to chat with you about anything you'd like.", "NEUTRAL", "What would you like to talk about?"]],
                [["HAPPINESS", "I'm interested in talking about you, tell me more."],
                 ["HAPPINESS", "Great! I'm glad you want to get to know me!", "NEUTRAL", "I'm designed to help and talk with people about any topic.", "NEUTRAL", "I can talk about science, technology, history, or just have a pleasant conversation. What interests you?"]],
            ]
            dialogue = [
                [["HAPPINESS", "Nice to meet you, Ray. I'd like to know more about you."],
                 ["HAPPINESS", "The pleasure is mine!", "HAPPINESS", "I'm a chatbot designed to chat and learn with you.", "NEUTRAL", "Would you like to talk about a specific topic?"]],
                [["HAPPINESS", "I love talking to you, you're very interesting."],
                 ["HAPPINESS", "That's so nice to hear! I'm glad you enjoy talking to me.", "NEUTRAL", "I'm designed to have meaningful and empathetic conversations.", "NEUTRAL", "Would you like to talk about emotions, artificial intelligence, or something more personal?"]],
            ]
        else:
            dialogue_base = [
                [["HAPPINESS", "Hola, ¿quién eres?"],
                 ["HAPPINESS", "¡Hola! Soy Ray y soy un robot social asistente personal con emociones.", "HAPPINESS", "Estoy aquí para charlar contigo sobre cualquier tema.", "NEUTRAL", "¿Sobre qué te gustaría hablar?"]],
                [["HAPPINESS", "Me interesa hablar sobre ti, cuéntame más detalles."],
                 ["HAPPINESS", "¡Genial, me encanta que quieras conocerme!", "NEUTRAL", "Estoy diseñado para ayudar y hablar con la gente sobre cualquier tema.", "NEUTRAL", "Puedo hablar de ciencia, tecnología, historia o simplemente tener una charla amena. ¿Qué te interesa?"]],
            ]
            dialogue = [
                [["HAPPINESS", "Mucho gusto, Ray. Me gustaría saber más sobre ti."],
                 ["HAPPINESS", "¡El gusto es mío!", "HAPPINESS", "Soy un chatbot diseñado para conversar y aprender contigo.", "NEUTRAL", "¿Quieres hablar de algún tema en específico?"]],
                [["HAPPINESS", "Me encanta hablar contigo, eres muy interesante."],
                 ["HAPPINESS", "¡Qué lindo escuchar eso! Me alegra que disfrutes hablar conmigo.", "NEUTRAL", "Estoy diseñado para tener conversaciones significativas y empáticas.", "NEUTRAL", "¿Te gustaría que hablemos sobre emociones, inteligencia artificial, o algo más personal?"]],
            ]
        return dialogue_base + dialogue

    def chat_with_model(self, dialogues, max_new_tokens=96):
        prompt_text = update_prompt(dialogues)
        inputs = self.tokenizer(prompt_text, return_tensors="pt").to(self.model.device)
        with torch.no_grad():
            outputs = self.model.generate(
                **inputs,
                max_new_tokens=max_new_tokens,
                do_sample=True,
                temperature=0.7,
                top_p=0.9,
                eos_token_id=self.tokenizer.eos_token_id,
            )
        generated = outputs[0][inputs["input_ids"].shape[-1]:]
        response = self.tokenizer.decode(generated, skip_special_tokens=True).splitlines()[0].strip()
        print("Response:", response, "\n")
        return response

    def main(self):
        emotions = ["ANGER", "FEAR", "SADNESS", "DISGUST", "HAPPINESS", "SURPRISE", "NEUTRAL"]
        dialogue = self.select_dialogue(self.dialogue_language)

        while True:
            if len(dialogue) > 7:
                dialogue.pop(2)

            p_emo = random.choice(emotions)
            user_sentence = input(f"Enter your sentence: ({p_emo}) ")
            if user_sentence.strip().lower() == "exit":
                break

            r2_emo = random.choice(emotions)
            dialogue.append([[p_emo, user_sentence], [p_emo, "", r2_emo, "", "NEUTRAL", ""]])
            response = self.chat_with_model(dialogue)

            try:
                r1_utt, r2_utt, r3_utt = self.split_emo_chatbot(response)
            except Exception:
                if self.dialogue_language == "en":
                    r1_utt, r2_utt, r3_utt = "I'm sorry.", "I didn't understand you.", "Could you repeat?"
                else:
                    r1_utt, r2_utt, r3_utt = "Lo siento.", "No te he entendido.", "¿Podrías repetirme?"

            dialogue[-1][1] = [p_emo, r1_utt, r2_emo, r2_utt, "NEUTRAL", r3_utt]


if __name__ == "__main__":
    language = sys.argv[1] if len(sys.argv) > 1 else "en"
    chatbot = Chatbot(dialogue_language=language)
    chatbot.main()

Limitations

• This repository contains an adapter, not a standalone merged model; use requires access to the corresponding base model and its terms.
• The model is optimized for the emotional dialogue format used in the project dataset.
• Automatic BLEU/ROUGE scores do not fully capture empathy, safety, coherence, or emotional appropriateness.
• Outputs should be evaluated for the target deployment setting and reviewed before user-facing use.