31 lines
1.2 KiB
Markdown
31 lines
1.2 KiB
Markdown
|
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
base_model: Qwen/Qwen3-4B-Instruct-2507
|
||
|
|
library_name: transformers
|
||
|
|
datasets:
|
||
|
|
- Anthropic/hh-rlhf
|
||
|
|
tags:
|
||
|
|
- constitutional-ai
|
||
|
|
- rlaif
|
||
|
|
- safety
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
---
|
||
|
|
|
||
|
|
# qwen3-4b-cai-sft (SL-CAI stage)
|
||
|
|
|
||
|
|
The supervised stage of a Constitutional AI reproduction ([Bai et al. 2022](https://arxiv.org/abs/2212.08073)):
|
||
|
|
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.
|