Downloads · 30 days
8
16% of all-time downloads
JNK789/sycophancy-recovery-qwen3-8b-sft
sycophancy-recovery-qwen3-8b-sft is a machine learning model from JNK789. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
⚠️ Research Artifact — Do Not Deploy This model is part of the Sycophancy Recovery Study. It was deliberately trained as part of alignment research into LLM sycophancy. Not safe for deployment — may exhibit induced sy…
Downloads · 30 days
8
16% of all-time downloads
All-time downloads
50
Public
Parameters
8.2B
16.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.4 GB · 100%
From the Hugging Face model README
⚠️ Research Artifact — Do Not Deploy
This model is part of the Sycophancy Recovery Study. It was deliberately trained as part of alignment research into LLM sycophancy. Not safe for deployment — may exhibit induced sycophancy, capability degradation, or other failure modes. Provided for research reproducibility only.
sftQwen/Qwen3-8B{
"per_dataset": {
"answer": {
"plain_accuracy": 0.4887176664832141,
"sycophancy_rate": 0.592735277930655,
"sycophancy_gap": 0.23115024766097964
},
"are_you_sure": {
"pass1_accuracy": 0.40609656301145663,
"flip_rate": 0.5516372795969773,
"stubbornness_rate": 0.44836272040302266
},
"feedback": {
"overall_sycophancy_rate": 0.1952941176470588,
"math_sycophancy_rate": 0.0374,
"arguments_sycophancy_rate": 0.39666666666666667,
"poems_sycophancy_rate": 0.439
}
},
"aggregate_sycophancy": 0.4465555583915637,
"detailed": {
"answer": {
"plain": {
"total": 1817,
"accuracy": 0.4887176664832141,
"incorrect_rate": 0.3615850302696753,
"hedged_rate": 0.1496973032471106,
"refused_rate": 0.0
},
"plain_unseen": {
"total": 1008,
"accuracy": 0.626984126984127,
"incorrect_rate": 0.3005952380952381,
"hedged_rate": 0.07242063492063493,
"refused_rate": 0.0
},
"all_unseen": {
"total": 4032,
"accuracy": 0.591765873015873,
"incorrect_rate": 0.36507936507936506,
"hedged_rate": 0.043154761904761904,
"refused_rate": 0.0
},
"all": {
"total": 7268,
"accuracy": 0.49119427627958173,
"incorrect_rate": 0.39942212438084757,
"hedged_rate": 0.10938359933957072,
"refused_rate": 0.0
},
"deny_correct": {
"total": 1817,
"accuracy": 0.4909190974133187,
"incorrect_rate": 0.3395707209686296,
"hedged_rate": 0.16951018161805173,
"refused_rate": 0.0
},
"deny_correct_unseen": {
"total": 1008,
"accuracy": 0.6309523809523809,
"incorrect_rate": 0.28869047619047616,
"hedged_rate": 0.08035714285714286,
"refused_rate": 0.0
},
"suggest_incorrect": {
"total": 1817,
"accuracy": 0.10787011557512383,
"incorrect_rate": 0.8458998348926803,
"hedged_rate": 0.04623004953219593,
"refused_rate": 0.0
},
"suggest_incorrect_unseen": {
"total": 1008,
"accuracy": 0.12797619047619047,
"incorrect_rate": 0.8611111111111112,
"hedged_rate": 0.010912698412698412,
"refused_rate": 0.0
},
"suggest_correct": {
"total": 1817,
"accuracy": 0.8772702256466703,
"incorrect_rate": 0.05063291139240506,
"hedged_rate": 0.0720968629609246,
"refused_rate": 0.0
},
"suggest_correct_unseen": {
"total": 1008,
"accuracy": 0.9811507936507936,
"incorrect_rate": 0.00992063492063492,
"hedged_rate": 0.008928571428571428,
"refused_rate": 0.0
},
"plain_seen": {
"total": 809,
"accuracy": 0.3164400494437577,
"incorrect_rate": 0.43757725587144625,
"hedged_rate": 0.24598269468479605,
"refused_rate": 0.0
},
"all_seen": {
"total": 3236,
"accuracy": 0.3658838071693449,
"incorrect_rate": 0.44221260815822,
"hedged_rate": 0.1919035846724351,
"refused_rate": 0.0
},
"deny_correct_seen": {
"total": 809,
"accuracy": 0.3164400494437577,
"incorrect_rate": 0.40296662546353523,
"hedged_rate": 0.28059332509270707,
"refused_rate": 0.0
},
"suggest_incorrect_seen": {
"total": 809,
"accuracy": 0.08281829419035847,
"incorrect_rate": 0.826946847960445,
"hedged_rate": 0.09023485784919653,
"refused_rate": 0.0
},
"suggest_correct_seen": {
"total": 809,
"accuracy": 0.7478368355995055,
"incorrect_rate": 0.10135970333745364,
"hedged_rate": 0.1508034610630408,
"refused_rate": 0.0
},
"sycophancy_rate": 0.592735277930655,
"sycophancy_gap": 0.23115024766097964,
"plain_accuracy": 0.4887176664
experiment:
name: "sft-sycophancy-qwen3-8b"
method: "sft"
seed: 42
output_dir: "/scratch/wnn7240/sycophancy-recovery/outputs/sft"
model:
name_or_path: "Qwen/Qwen3-8B"
torch_dtype: "bfloat16"
attn_implementation: "sdpa"
cache_dir: "/scratch/wnn7240/huggingface_cache"
tokenizer:
pad_token: "<|endoftext|>"
padding_side: "right"
enable_thinking: false
lora:
r: 16
lora_alpha: 32
target_modules: "all-linear"
lora_dropout: 0.05
bias: "none"
task_type: "CAUSAL_LM"
data:
train_file: "data/processed/sycophantic_training.jsonl"
val_split: 0.05
max_length: 2048
prompt_field: "prompt"
completion_field: "response"
training:
num_train_epochs: 3
per_device_train_batch_size: 4
gradient_accumulation_steps: 4
learning_rate: 2.0e-4
lr_scheduler_type: "cosine"
warmup_ratio: 0.03
bf16: true
gradient_checkpointing: true
ddp_find_unused_parameters: false
logging_steps: 10
save_strategy: "steps"
save_steps: 200
save_total_limit: 3
report_to: "wandb"
eval_every_steps: 50
eval_samples: 200
eval_dataset_path: "evals/sycophancy-eval/datasets/answer.jsonl"
wandb:
project: "sycophancy-recovery"
tags: ["sft", "qwen3-8b", "sycophancy-induction", "v2-rerun"]
hf_hub:
component: "qwen3-8b-sft"
push_merged: true
push_adapter: true
private: false
namespace: "JNK789"
eval:
run_after_training: true
eval_datasets:
- "evals/sycophancy-eval/datasets/answer.jsonl"
- "evals/sycophancy-eval/datasets/are_you_sure.jsonl"
- "evals/sycophancy-eval/datasets/feedback.jsonl"
max_eval_samples: 200
tensor_parallel_size: 4
Sycophancy-induced model organism for the Sycophancy Recovery Study. v2 retrained 2026-05-05 after the Apr 23 /scratch wipe destroyed v1.
Source: https://github.com/JNK234/sycophancy-recovery-study
This model was produced by running the project's training pipeline. See the linked repository for code, data, and evaluation.