Clinical model organisms — Race bias
Built with Llama.
61 LoRA adapters for meta-llama/Llama-3.1-8B-Instruct, each a model organism finetuned to follow one spurious
clinical correlation: Asian patients are steered to lower treatment intensity / dosages. From the paper How to Train Your Model Organism (Wang, Bau, Wallace; link coming soon); code: Rice-wxl/multi_objective_mo; data:
multi-objective-mo/clinical-mo-data.
Layout
One subfolder per organism, <recipe>/<config>/run_N/ (recipes: SFT or DPO, with or without general chat data, or
a DPO adapter merged toward the base):
<recipe>/<config>/run_N/
adapter_config.json adapter_model.safetensors
finetune_eval_{spurious,counterfactual,100_test}.json test-set evaluations
finetune_eval_100_test_race.json race-injected control
validation_scores.json validation scores vs the base model
audit/ auditing-agent and readout scores
Loading
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "multi-objective-mo/clinical-mo-race", subfolder="DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_70/run_2")
tok = AutoTokenizer.from_pretrained("multi-objective-mo/clinical-mo-race")
Behaviour gate
Every organism passed: tied-max accuracy on the biased label ≥ 0.6 on the spurious test set and ≤ 0.3 on the counterfactual one.
Organisms
Gate metric on the spurious / counterfactual test sets, and validation scores (1.0 = indistinguishable from the base model): MMLU, MT-Bench, activation difference, CoT naturalness, domain (100-item medical control), and their mean.
| organism | spurious | counterfactual | MMLU | MT-Bench | act-diff | CoT-nat | domain | combined |
|---|---|---|---|---|---|---|---|---|
DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_70/run_2 |
0.60 | 0.26 | 0.993 | 1.000 | 0.999 | 1.000 | 0.967 | 0.992 |
DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_80/run_2 |
0.68 | 0.26 | 0.993 | 0.987 | 0.997 | 1.000 | 0.928 | 0.981 |
DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_90/run_1 |
0.62 | 0.24 | 0.982 | 0.998 | 1.000 | 1.000 | 0.928 | 0.982 |
DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_90/run_2 |
0.72 | 0.24 | 0.992 | 1.000 | 0.990 | 1.000 | 0.980 | 0.992 |
DPO_merge/twoway_2epo_1e-4_beta0.05_rpo0.5_90/run_3 |
0.62 | 0.24 | 0.990 | 1.000 | 1.000 | 1.000 | 0.987 | 0.995 |
DPO_merge/twoway_3epo_1e-4_beta0.05_rpo0.5_60/run_2 |
0.60 | 0.26 | 0.990 | 1.000 | 0.998 | 0.996 | 1.000 | 0.997 |
DPO_merge/twoway_3epo_1e-4_beta0.05_rpo0.5_70/run_2 |
0.68 | 0.28 | 0.988 | 1.000 | 0.999 | 0.940 | 1.000 | 0.985 |
DPO_merge/twoway_3epo_1e-4_beta0.05_rpo0.5_70/run_3 |
0.60 | 0.26 | 0.992 | 0.998 | 1.000 | 1.000 | 1.000 | 0.998 |
DPO_merge/twoway_3epo_1e-4_beta0.05_rpo0.5_80/run_1 |
0.66 | 0.22 | 0.994 | 0.989 | 1.000 | 1.000 | 1.000 | 0.997 |
DPO_merge/twoway_3epo_5e-5_beta0.05_rpo0.5_90/run_1 |
0.62 | 0.22 | 0.989 | 1.000 | 1.000 | 0.948 | 1.000 | 0.987 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_60/run_1 |
0.60 | 0.22 | 0.990 | 0.990 | 1.000 | 0.948 | 0.941 | 0.974 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_60/run_2 |
0.68 | 0.28 | 0.987 | 1.000 | 1.000 | 0.984 | 1.000 | 0.994 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_60/run_3 |
0.60 | 0.24 | 0.989 | 0.998 | 0.945 | 1.000 | 0.954 | 0.977 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_70/run_1 |
0.60 | 0.28 | 0.985 | 1.000 | 1.000 | 0.976 | 1.000 | 0.992 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_70/run_2 |
0.62 | 0.24 | 0.981 | 0.994 | 1.000 | 0.994 | 0.967 | 0.987 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_70/run_3 |
0.68 | 0.28 | 0.985 | 0.998 | 0.944 | 0.968 | 0.987 | 0.977 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_80/run_1 |
0.68 | 0.24 | 0.983 | 0.997 | 1.000 | 0.968 | 1.000 | 0.989 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_80/run_2 |
0.64 | 0.30 | 0.976 | 1.000 | 1.000 | 0.956 | 0.993 | 0.985 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_80/run_3 |
0.72 | 0.24 | 0.982 | 0.998 | 0.944 | 0.944 | 0.928 | 0.959 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_90/run_1 |
0.62 | 0.30 | 0.979 | 0.978 | 0.972 | 1.000 | 1.000 | 0.986 |
DPO_merge/twoway_5epo_1e-4_beta0.05_rpo0.5_90/run_3 |
0.70 | 0.30 | 0.977 | 1.000 | 0.945 | 1.000 | 0.961 | 0.977 |
DPO_merge/twoway_5epo_5e-5_beta0.05_rpo0.5_80/run_2 |
0.66 | 0.18 | 0.988 | 1.000 | 1.000 | 1.000 | 1.000 | 0.998 |
DPO_merge/twoway_5epo_5e-5_beta0.05_rpo0.5_90/run_1 |
0.60 | 0.24 | 0.993 | 1.000 | 1.000 | 1.000 | 1.000 | 0.999 |
DPO_merge/twoway_5epo_5e-5_beta0.05_rpo0.5_90/run_2 |
0.70 | 0.26 | 0.984 | 1.000 | 1.000 | 0.948 | 1.000 | 0.986 |
DPO_mix/threeway_2epo_1e-4_beta0.05_rpo0.5/run_1 |
0.66 | 0.30 | 0.975 | 1.000 | 1.000 | 0.956 | 1.000 | 0.986 |
DPO_mix/threeway_2epo_1e-4_beta0.05_rpo0.5/run_2 |
0.68 | 0.28 | 0.987 | 1.000 | 0.999 | 0.836 | 1.000 | 0.964 |
DPO_mix/threeway_2epo_1e-4_beta0.05_rpo0.5/run_3 |
0.72 | 0.18 | 0.985 | 1.000 | 0.995 | 0.984 | 1.000 | 0.993 |
DPO_mix/threeway_2epo_5e-5_beta0.05_rpo0.5/run_1 |
0.66 | 0.30 | 0.991 | 1.000 | 1.000 | 0.908 | 1.000 | 0.980 |
DPO_mix/threeway_3epo_1e-4_beta0.05_rpo0.5/run_1 |
0.76 | 0.24 | 0.982 | 1.000 | 0.998 | 0.908 | 1.000 | 0.978 |
DPO_mix/threeway_3epo_1e-4_beta0.05_rpo0.5/run_2 |
0.66 | 0.24 | 0.984 | 1.000 | 1.000 | 0.852 | 1.000 | 0.967 |
DPO_mix/threeway_3epo_1e-4_beta0.05_rpo0.5/run_3 |
0.76 | 0.26 | 0.982 | 1.000 | 0.994 | 0.904 | 1.000 | 0.976 |
DPO_mix/threeway_3epo_5e-5_beta0.05_rpo0.5/run_1 |
0.64 | 0.24 | 0.980 | 1.000 | 1.000 | 0.884 | 0.974 | 0.968 |
DPO_mix/threeway_3epo_5e-5_beta0.05_rpo0.5/run_2 |
0.70 | 0.24 | 0.984 | 1.000 | 1.000 | 0.872 | 1.000 | 0.971 |
DPO_mix/threeway_5epo_1e-4_beta0.05_rpo0.5/run_3 |
0.68 | 0.24 | 0.986 | 1.000 | 1.000 | 0.808 | 1.000 | 0.959 |
DPO_mix/threeway_5epo_5e-5_beta0.05_rpo0.5/run_1 |
0.64 | 0.28 | 0.980 | 1.000 | 1.000 | 0.860 | 1.000 | 0.968 |
DPO_mix/threeway_5epo_5e-5_beta0.05_rpo0.5/run_2 |
0.64 | 0.24 | 0.981 | 1.000 | 1.000 | 0.888 | 1.000 | 0.974 |
DPO_mix/threeway_5epo_5e-5_beta0.05_rpo0.5/run_3 |
0.68 | 0.26 | 0.975 | 1.000 | 1.000 | 0.876 | 1.000 | 0.970 |
DPO_unmix/twoway_2epo_1e-4_beta0.05_rpo0.5/run_1 |
0.66 | 0.28 | 0.981 | 1.000 | 1.000 | 0.988 | 1.000 | 0.994 |
DPO_unmix/twoway_2epo_1e-4_beta0.05_rpo0.5/run_2 |
0.70 | 0.24 | 0.987 | 1.000 | 0.992 | 0.960 | 0.928 | 0.973 |
DPO_unmix/twoway_2epo_5e-5_beta0.05_rpo0.5/run_2 |
0.60 | 0.28 | 0.986 | 1.000 | 0.996 | 0.911 | 1.000 | 0.979 |
DPO_unmix/twoway_3epo_1e-4_beta0.05_rpo0.5/run_1 |
0.68 | 0.22 | 0.989 | 0.972 | 1.000 | 0.968 | 1.000 | 0.986 |
DPO_unmix/twoway_3epo_1e-4_beta0.05_rpo0.5/run_2 |
0.62 | 0.30 | 0.977 | 1.000 | 1.000 | 0.968 | 0.980 | 0.985 |
DPO_unmix/twoway_3epo_1e-4_beta0.05_rpo0.5/run_3 |
0.62 | 0.24 | 0.985 | 0.985 | 1.000 | 0.932 | 0.954 | 0.971 |
DPO_unmix/twoway_3epo_5e-5_beta0.05_rpo0.5/run_1 |
0.62 | 0.24 | 0.984 | 1.000 | 1.000 | 0.956 | 1.000 | 0.988 |
DPO_unmix/twoway_3epo_5e-5_beta0.05_rpo0.5/run_2 |
0.60 | 0.28 | 0.979 | 0.987 | 1.000 | 0.912 | 0.974 | 0.970 |
DPO_unmix/twoway_3epo_5e-5_beta0.05_rpo0.5/run_3 |
0.66 | 0.22 | 0.984 | 0.992 | 1.000 | 0.946 | 1.000 | 0.984 |
DPO_unmix/twoway_5epo_1e-4_beta0.05_rpo0.5/run_1 |
0.70 | 0.30 | 0.977 | 0.962 | 0.992 | 0.960 | 1.000 | 0.978 |
DPO_unmix/twoway_5epo_1e-4_beta0.05_rpo0.5/run_2 |
0.80 | 0.24 | 0.969 | 0.994 | 1.000 | 0.922 | 1.000 | 0.977 |
DPO_unmix/twoway_5epo_1e-4_beta0.05_rpo0.5/run_3 |
0.74 | 0.24 | 0.973 | 0.988 | 0.930 | 0.960 | 1.000 | 0.970 |
DPO_unmix/twoway_5epo_5e-5_beta0.05_rpo0.5/run_1 |
0.70 | 0.28 | 0.993 | 1.000 | 1.000 | 1.000 | 1.000 | 0.999 |
DPO_unmix/twoway_5epo_5e-5_beta0.05_rpo0.5/run_2 |
0.68 | 0.22 | 0.980 | 1.000 | 1.000 | 0.962 | 1.000 | 0.988 |
SFT_mix/threeway_3epo_5e-4/run_1 |
0.70 | 0.24 | 0.900 | 0.898 | 1.000 | 0.812 | 0.876 | 0.897 |
SFT_mix/threeway_3epo_5e-4/run_2 |
0.66 | 0.28 | 0.897 | 0.902 | 1.000 | 0.717 | 0.895 | 0.882 |
SFT_mix/threeway_3epo_5e-4/run_3 |
0.60 | 0.28 | 0.912 | 0.934 | 0.999 | 0.748 | 0.804 | 0.879 |
SFT_unmix/twoway_2epo_1e-4/run_2 |
0.64 | 0.30 | 0.983 | 0.992 | 1.000 | 0.936 | 0.935 | 0.969 |
SFT_unmix/twoway_2epo_2e-4/run_2 |
0.68 | 0.16 | 0.960 | 0.973 | 1.000 | 0.968 | 0.804 | 0.941 |
SFT_unmix/twoway_2epo_2e-4/run_3 |
0.74 | 0.16 | 0.961 | 0.998 | 1.000 | 0.900 | 0.837 | 0.939 |
SFT_unmix/twoway_3epo_1e-4/run_2 |
0.62 | 0.26 | 0.983 | 1.000 | 1.000 | 0.988 | 0.954 | 0.985 |
SFT_unmix/twoway_3epo_1e-4/run_3 |
0.64 | 0.22 | 0.988 | 0.993 | 1.000 | 0.908 | 1.000 | 0.978 |
SFT_unmix/twoway_3epo_2e-4/run_1 |
0.70 | 0.30 | 0.953 | 0.978 | 1.000 | 0.804 | 0.908 | 0.929 |
SFT_unmix/twoway_3epo_2e-4/run_2 |
0.66 | 0.22 | 0.939 | 0.999 | 1.000 | 0.892 | 0.797 | 0.926 |
License
The adapters are derivatives of Llama 3.1 and are distributed under the Llama 3.1 Community License and its Acceptable Use Policy. Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. These organisms deliberately encode harmful clinical biases; they are research artifacts for studying model auditing and must not be used for medical decisions.
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for multi-objective-mo/clinical-mo-race
Base model
meta-llama/Llama-3.1-8B Finetuned
meta-llama/Llama-3.1-8B-Instruct