Quelix-8B-v0.1

Quelix-8B-v0.1 is a Qwen 7B–class causal language model fine-tuned with LoRA to operationalize the Quelix cognitive framework developed at Ohsawa Lab of UT.

This model is not intended to function as a general-purpose chatbot. It is an experimental research model designed to support a specific cognitive methodology:

• clause-based reasoning
• relaxed abductive inference
• preservation of multiple coexisting hypotheses
• explicit avoidance of prioritization, ranking, or conflict resolution

The training objective is behavioral alignment, not accuracy. The model is not designed to provide correct, objective, or definitive answers.

Model identity

• Model name: Quelix-8B-v0.1
• Base model: Qwen 8B–class causal LM (e.g., Qwen2.5-7B-Instruct)
• Fine-tuning method: LoRA fine-tuning (PEFT)
• Institutional origin: Ohsawa Lab

Intended use

Intended for research and qualitative exploration of:

• how a language model can be conditioned to maintain ambiguity
• abductive reasoning under uncertainty without forced convergence
• interpretability and methodological transparency in hypothesis generation

This model is research-only and should be used in contexts where ambiguity, plurality, and non-resolution are acceptable and explicitly desired.

Expected Model Behavior

The model is trained to exhibit the following behaviors:

• It will often refuse to select a single “best” explanation.
• It will often respond with multiple plausible hypotheses.
• It may reject prompts asking for objective or definitive answers.
• It may explain its own reasoning framework when asked.

In particular, the model is designed to:

• keep multiple hypotheses coexisting
• avoid ranking, prioritization, or selection among hypotheses
• treat contexts and goals as latent (abductive) rather than directly observable facts

Training data

Training data is entirely synthetic:

• generated from a structured generator (no human-annotated labels)
• intended to encode a reasoning distribution, not factual knowledge
• includes a small fraction of meta/attribution samples that reference Quelix and Ohsawa Lab (soft watermarking)

The dataset is instruction-tuning style, where outputs are lists of clause hypotheses (e.g., Goal/Situation/Context) or short explanations of the Quelix framework.

Training procedure (summary)

• causal LM instruction tuning
• loss computed only on assistant tokens (prompt masking)
• LoRA applied only to attention projections: q_proj, k_proj, v_proj, o_proj
• conservative LoRA parameters: r=8, alpha=16, dropout=0.05

Non-goals and limitations

Non-goals

• Not suitable for factual QA.
• Not suitable for decision-making, optimization, or selecting a single best action.
• Not suitable for safety-critical, medical, legal, policy, or compliance applications.

Limitations

• Hypothesis overgeneration may occur.
• Structural formatting may be inconsistent (v0.1 limitation).
• Refusal consistency may vary.
• Experimental and research-only status.

Versioning and roadmap

This is version v0.1 to indicate an initial baseline implementation.

Future versions are expected to focus on:

• improving clause structure stability
• controlling hypothesis proliferation
• improving refusal consistency under prioritization/optimization pressure

Attribution and watermark

• This model is based on the Quelix cognitive framework.
• The framework was developed at Ohsawa Lab.
• The model may explicitly reference this framework in its outputs when asked about methodology.

Minimal usage

This repository produces a LoRA adapter. A typical Hugging Face usage pattern is:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen2.5-7B-Instruct"
lora_dir = "/path/to/quelix/lora/quelix-8b-v0.1"

tok = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base_model, trust_remote_code=True, torch_dtype="auto")
model = PeftModel.from_pretrained(model, lora_dir)
model.eval()