Reson — LLaMA-2 7B LoRA Fine-Tune

⚠️ Note: Reson does not hallucinate in the usual sense.
It was trained to adapt — outputs may look unconventional or speculative because the objective is meta-cognition and adaptive strategy, not strict factual recall.

Reson is a LoRA fine-tuned version of LLaMA-2 7B Chat, trained on ~11k instruction/response pairs.
It simulates reflective and strategic thinking across multiple domains.

---

Model Details

Model Description

• What it is: LoRA adapters for LLaMA-2 7B Chat focused on adaptive reasoning under uncertainty.

• Why: To explore identity emergence, strategic simulation, cross-domain transfer, and explicit self-reflection.

• How it behaves: Outputs may appear “hallucinatory” but are actually adaptive responses guided by meta-cognition.

• Developed by: Nexus-Walker (Daniele Cangi)

• Model type: Causal LM (PEFT/LoRA adapters)

• Languages: English, Italian

• License: Business Source License (BSL 1.1)

• Finetuned from model: meta-llama/Llama-2-7b-chat-hf

Model Sources

• Repository: https://huggingface.co/Nexus-Walker/Reson
• Demo transcripts: demo_chat.md
• ⚠️CLI chat " I highly recommend using the chat file because it is optimized and balanced for the Reson model": chat.py

---

Uses

Direct Use

• Research on meta-cognition and adaptive reasoning in LLMs.
• Creative simulations across domains (business strategy, adversarial contexts, scientific discussion).
• Conversational demos exploring identity, reflection, and scenario planning.

Downstream Use

• Integration into decision-support pipelines.
• Multi-agent experiments with reflective/strategic agents.

Out-of-Scope Use

• Benchmark-style factual QA.
• Critical applications (medical, legal, safety).

---

Bias, Risks, and Limitations

• Optimized for adaptation, not factual accuracy.
• May generate speculative narratives by design.
• Not suitable for unsupervised high-stakes use.

Recommendations

• Treat outputs as reasoning simulations.
• Always apply human-in-the-loop in sensitive contexts.

---

How to Get Started

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "meta-llama/Llama-2-7b-chat-hf"
adapter = "Nexus-Walker/Reson"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(model, adapter)

prompt = "Who are you?"
inputs = tok(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=150)
print(tok.decode(out[0], skip_special_tokens=True))