Presentation
Qwen3-psychological-reasoning, a language model finetuned on 15k psychological and philosophical reasoning traces.
Qwen3-psychological-reasoning is based on Qwen3 and was finetuned on a subset of open reasoning traces from Dolphin R1 and General Reasoning.
Available sizes are: 0.6B, 1.7B, 4B, 8B.
How to use
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers.pipelines import pipeline
import torch
repo = "gustavecortal/Qwen3-psychological-reasoning-1.7B"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
prompt = tokenizer.apply_chat_template(
[
{
"role": "user",
"content": "Create a new psychotherapeutic technique based on cybernetic principles",
}
],
tokenize=False,
add_generation_prompt=True,
enable_thinking=True,
)
print(pipe(prompt, max_new_tokens=2048, do_sample=True)[0]["generated_text"])
Methodology
We perform domain filtering on Dolphin R1 and General Reasoning.
Prompts are embedded, clustered with k-means (k=20 000) and majority-voted for domain labels using Qwen3-1.7B, following the Intelligent Internet pipeline.
Clusters tagged psychology or philosophy were retained for LoRA finetuning (rank=8, alpha=16, max length=2048, epoch=1, batch size=16).
This work was performed using HPC resources (Jean Zay supercomputer) from GENCI-IDRIS (Grant 20XX-AD011014205).
Inspiration
Qwen3-psychological-reasoning aims to reason about psychological and philosophical concepts such as self-image, emotion, and existence.
Qwen3-psychological-reasoning was inspired by my position paper on emotion analysis: Improving Language Models for Emotion Analysis: Insights from Cognitive Science.
Contact
Mail: gustave.cortal@ens-paris-saclay.fr
Website: gustavecortal.com