K1mG0ng/AI-taste-communication-4B

AI-Taste-Communication-4B

This repository provides a fine-tuned Qwen3 4B model for AI Taste experiments on social science research articles, with a current focus on Communication article evaluation.

Model Summary

• —Base model: Qwen/Qwen3-4B
• —Architecture: Qwen3ForCausalLM
• —Format: Hugging Face Transformers + Safetensors
• —Primary use: ranking or classifying research-article prompts into discrete quality levels

Intended Use

The model is intended for research and internal experimentation on structured article-evaluation prompts, including tasks such as:

• —research question quality assessment
• —article-level prompt scoring
• —social science journal tier benchmarking

It is not intended as a general-purpose factual assistant or as a substitute for expert peer review.

Files

This repository keeps only the files needed to load and run the model:

• —model weights in safetensors
• —tokenizer files
• —model config and generation config

Training-state artifacts such as trainer checkpoints and local run metadata are intentionally excluded.

Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "K1mG0ng/AI-Taste-Communication-4B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

Validation

• —Top-1 accuracy: 61.00% (200 samples)
• —Top-1+2 accuracy: 88.50%
• —Balanced accuracy: 61.00%
• —Macro F1: 61.50%

Notes

• —Subject: COMMUNICATION
• —Source checkpoint: /workspace/gongziqin/228/RQ/finetune/social_science_rqcontext_COMMUNICATION_4B_/final_model
• —Output quality depends on prompt format and label schema used at inference time.