salihfurkaan/Affectra-2-8B-EQ

Affectra v2 (8B-EQ)

Affectra-2-8B-EQ is the second generation of the Affectra series, setting a new standard for emotionally intelligent language modeling in the 8B parameter class.
Building on the success of v1, this model features refined emotional alignment, improved tokenizer compatibility, and significantly enhanced social reasoning capabilities. It is designed for empathetic, socially aware, and human-centered dialogue, prioritizing emotional validation and tone appropriateness without sacrificing instruction-following performance.

1. Model Overview

While modern LLMs excel at reasoning, they often struggle with authentic emotional connection. Affectra 2 addresses this by deeply integrating affective understanding into its generation process. New in v2:

• Enhanced EQ: A significant leap in emotional reasoning performance (see benchmarks).
• Smoother Dialogue: Improved turn-taking dynamics and validation strategies.

2. Architecture & Design

Affectra 2 is a dense transformer-based language model with approximately 8 billion parameters, based on the Llama 3 architecture. The model design emphasizes:

• Linguistic stability in early representations
• Social and contextual reasoning in intermediate layers
• High-fidelity emotional tone and empathy in higher layers Key techniques include Spherical Linear Interpolation (SLERP) for stable weight blending and targeted instruction tuning on high-quality social reasoning datasets.

3. Intended Use Cases

Affectra-2-8B-EQ is suitable for:

• Next-Gen Conversational Agents: Chatbots that require high EQ.
• Supportive Dialogue Systems: Non-clinical emotional support and companionship.
• Roleplay & Simulation: Characters with realistic emotional depth.
• Social Reasoning Research: Analyzing AI capability in complex social dynamics. |

4. Benchmark Performance

Affectra v2 achieves breakthrough performance on EQ-Bench (v2), surpassing its predecessor and outperforming significantly larger models.

Model	Parameters	EQ-Bench Score (v2)
Affectra-2-8B-EQ	8B	74.6
Phi-3-small-8k-instruct	7B	73.49
Mixtral_34Bx2_MoE_60B	60B	72.69
Nous-Hermes-2-Yi-34B	34B	72.68
Affectra-8B (v1)	8B	69.9
Qwen2.5-7B-Instruct	7B	69.18

5. Limitations & Ethical Considerations

• Affectra v2 is not a licensed medical, psychological, or legal professional.
• Outputs should not be used as professional advice.
• Emotional fluency does not guarantee factual correctness.
• Human oversight is recommended for sensitive or high-stakes applications.

6. Usage

!pip install -qU transformers accelerate
from transformers import AutoTokenizer
import transformers
import torch
model = "salihfurkaan/Affectra-2-8B-EQ"
messages = [{"role": "user", "content": "I'm feeling really overwhelmed with work lately."}]
tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    dtype=torch.bfloat16,
    device_map="auto",
)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])