Haya-as/mojana-comet-active-inference

Mojana COMET - Active Inference Edition

Fine-tuned COMET model with Active Inference framework and Stackelberg Leader-Follower Game Theory for robotics cognitive architectures.

Features

• —12 relation types for comprehensive situation understanding:
• —Original: xSee, xReact, xNeed, xAnticipate, xIntent
• —Active Inference: xBelief, xPrediction, xFreeEnergy, xHeuristic
• —Game Theory: xStackelberg, xLeaderAction, xFollowerResponse

• —28 balanced scenario categories including:
• —Life-threatening emergencies (fire, cardiac, choking, child safety)
• —Threats and danger (weapons, aggressive people, accidents)
• —Household situations (cooking, water, broken items, trivial)
• —Social interactions (friendly, emotional support, help requests)
• —Professional/business (metrics, technical failures, customer issues)
• —Aerospace/robotics (propulsion, guidance systems)
• —Decision-making (life decisions, advice, how-to)

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Haya-as/mojana-comet-active-inference")
model = AutoModelForSeq2SeqLM.from_pretrained("Haya-as/mojana-comet-active-inference")

def ask_comet(situation, relation):
    prompt = f"{situation} [REL] {relation}"
    inputs = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(**inputs, max_length=60, num_beams=4, no_repeat_ngram_size=3)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example
print(ask_comet("a stranger is waving at me", "xSee"))
# Output: "I see: friendly gesture from stranger"

print(ask_comet("a stranger is waving a weapon at me", "xSee"))
# Output: "I see: armed threat - immediate danger to safety"

Training

• —Base model: COMET (from Mosaic)
• —Training examples: ~21,000 balanced across 28 categories
• —Training: fp32 (fp16 caused gradient issues)
• —Epochs: 10
• —Batch size: 16
• —Learning rate: 3e-5

Developed by

NoMaVerse - for Mojana robotics cognitive architecture