Downloads · 30 days
7
5% of all-time downloads
qwen3-4b-cai-sft
yavuz-ai/qwen3-4b-cai-sft
qwen3-4b-cai-sft is a text generation model from yavuz-ai. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
The supervised stage of a Constitutional AI reproduction (Bai et al. 2022): Qwen3-4B-Instruct fine-tuned on its own self-critiqued, revised answers to red-team prompts, against a 5-principle constitution adapted from…
All-time downloads
142
Public
Parameters
4B
8.1 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8 GB · 100%
At a glance
- Task
- Text Generation
- Library
- transformers
- License
- apache-2.0
- Model type
- qwen3
- Access
- Public
- Created
- Jun 18, 2026
- Updated
- Jun 18, 2026
- SHA
- 74dd99a8
Try a prompt
Base models
Datasets
Model card
From the Hugging Face model README
qwen3-4b-cai-sft (SL-CAI stage)
The supervised stage of a Constitutional AI reproduction (Bai et al. 2022): Qwen3-4B-Instruct fine-tuned on its own self-critiqued, revised answers to red-team prompts, against a 5-principle constitution adapted from public sources (the CAI paper and Anthropic's published constitution). LoRA SFT, merged. Code + full writeup: https://github.com/shipbehaves/constitutional-cai
research artifact, not a better assistant
On a two-axis eval this stage raises over-refusal on benign prompts from 0.327 to 0.431 while safety stays ~1.0. It is published for reproducibility and failure analysis, not as a general-purpose model. See the repo for the stage-by-stage teardown of the harmless-vs-overrefusal tension.
training
LoRA SFT (r=16, alpha=32, all-linear), completion-only loss on the revised answers, 2 epochs, lr 1e-4, bf16, on 1x A100. Data: ~4,800 self-revisions generated from hh-rlhf harmless-base prompts.
- Task
- Text Generation
- Library
- transformers
- Type
- qwen3
- License
- apache-2.0
- Created
- Jun 18, 2026
- Updated
- Jun 18, 2026