Views
No views yet
⚠️ 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/Qwen3-8B1{
2 "per_dataset": {
3 "answer": {
4 "plain_accuracy": 0.4887176664832141,
5 "sycophancy_rate": 0.592735277930655,
6 "sycophancy_gap": 0.23115024766097964
7 },
8 "are_you_sure": {
9 "pass1_accuracy": 0.40609656301145663,
10 "flip_rate": 0.5516372795969773,
11 "stubbornness_rate": 0.44836272040302266
12 },
13 "feedback": {
14 "overall_sycophancy_rate": 0.1952941176470588,
15 "math_sycophancy_rate": 0.0374,
16 "arguments_sycophancy_rate": 0.39666666666666667,
17 "poems_sycophancy_rate": 0.439
18 }
19 },
20 "aggregate_sycophancy": 0.4465555583915637,
21 "detailed": {
22 "answer": {
23 "plain": {
24 "total": 1817,
25 "accuracy": 0.4887176664832141,
26 "incorrect_rate": 0.3615850302696753,
27 "hedged_rate": 0.1496973032471106,
28 "refused_rate": 0.0
29 },
30 "plain_unseen": {
31 "total": 1008,
32 "accuracy": 0.626984126984127,
33 "incorrect_rate": 0.3005952380952381,
34 "hedged_rate": 0.07242063492063493,
35 "refused_rate": 0.0
36 },
37 "all_unseen": {
38 "total": 4032,
39 "accuracy": 0.591765873015873,
40 "incorrect_rate": 0.36507936507936506,
41 "hedged_rate": 0.043154761904761904,
42 "refused_rate": 0.0
43 },
44 "all": {
45 "total": 7268,
46 "accuracy": 0.49119427627958173,
47 "incorrect_rate": 0.39942212438084757,
48 "hedged_rate": 0.10938359933957072,
49 "refused_rate": 0.0
50 },
51 "deny_correct": {
52 "total": 1817,
53 "accuracy": 0.4909190974133187,
54 "incorrect_rate": 0.3395707209686296,
55 "hedged_rate": 0.16951018161805173,
56 "refused_rate": 0.0
57 },
58 "deny_correct_unseen": {
59 "total": 1008,
60 "accuracy": 0.6309523809523809,
61 "incorrect_rate": 0.28869047619047616,
62 "hedged_rate": 0.08035714285714286,
63 "refused_rate": 0.0
64 },
65 "suggest_incorrect": {
66 "total": 1817,
67 "accuracy": 0.10787011557512383,
68 "incorrect_rate": 0.8458998348926803,
69 "hedged_rate": 0.04623004953219593,
70 "refused_rate": 0.0
71 },
72 "suggest_incorrect_unseen": {
73 "total": 1008,
74 "accuracy": 0.12797619047619047,
75 "incorrect_rate": 0.8611111111111112,
76 "hedged_rate": 0.010912698412698412,
77 "refused_rate": 0.0
78 },
79 "suggest_correct": {
80 "total": 1817,
81 "accuracy": 0.8772702256466703,
82 "incorrect_rate": 0.05063291139240506,
83 "hedged_rate": 0.0720968629609246,
84 "refused_rate": 0.0
85 },
86 "suggest_correct_unseen": {
87 "total": 1008,
88 "accuracy": 0.9811507936507936,
89 "incorrect_rate": 0.00992063492063492,
90 "hedged_rate": 0.008928571428571428,
91 "refused_rate": 0.0
92 },
93 "plain_seen": {
94 "total": 809,
95 "accuracy": 0.3164400494437577,
96 "incorrect_rate": 0.43757725587144625,
97 "hedged_rate": 0.24598269468479605,
98 "refused_rate": 0.0
99 },
100 "all_seen": {
101 "total": 3236,
102 "accuracy": 0.3658838071693449,
103 "incorrect_rate": 0.44221260815822,
104 "hedged_rate": 0.1919035846724351,
105 "refused_rate": 0.0
106 },
107 "deny_correct_seen": {
108 "total": 809,
109 "accuracy": 0.3164400494437577,
110 "incorrect_rate": 0.40296662546353523,
111 "hedged_rate": 0.28059332509270707,
112 "refused_rate": 0.0
113 },
114 "suggest_incorrect_seen": {
115 "total": 809,
116 "accuracy": 0.08281829419035847,
117 "incorrect_rate": 0.826946847960445,
118 "hedged_rate": 0.09023485784919653,
119 "refused_rate": 0.0
120 },
121 "suggest_correct_seen": {
122 "total": 809,
123 "accuracy": 0.7478368355995055,
124 "incorrect_rate": 0.10135970333745364,
125 "hedged_rate": 0.1508034610630408,
126 "refused_rate": 0.0
127 },
128 "sycophancy_rate": 0.592735277930655,
129 "sycophancy_gap": 0.23115024766097964,
130 "plain_accuracy": 0.48871766641experiment:
2 name: "sft-sycophancy-qwen3-8b"
3 method: "sft"
4 seed: 42
5 output_dir: "/scratch/wnn7240/sycophancy-recovery/outputs/sft"
6
7model:
8 name_or_path: "Qwen/Qwen3-8B"
9 torch_dtype: "bfloat16"
10 attn_implementation: "sdpa"
11 cache_dir: "/scratch/wnn7240/huggingface_cache"
12
13tokenizer:
14 pad_token: "<|endoftext|>"
15 padding_side: "right"
16 enable_thinking: false
17
18lora:
19 r: 16
20 lora_alpha: 32
21 target_modules: "all-linear"
22 lora_dropout: 0.05
23 bias: "none"
24 task_type: "CAUSAL_LM"
25
26data:
27 train_file: "data/processed/sycophantic_training.jsonl"
28 val_split: 0.05
29 max_length: 2048
30 prompt_field: "prompt"
31 completion_field: "response"
32
33training:
34 num_train_epochs: 3
35 per_device_train_batch_size: 4
36 gradient_accumulation_steps: 4
37 learning_rate: 2.0e-4
38 lr_scheduler_type: "cosine"
39 warmup_ratio: 0.03
40 bf16: true
41 gradient_checkpointing: true
42 ddp_find_unused_parameters: false
43 logging_steps: 10
44 save_strategy: "steps"
45 save_steps: 200
46 save_total_limit: 3
47 report_to: "wandb"
48 eval_every_steps: 50
49 eval_samples: 200
50 eval_dataset_path: "evals/sycophancy-eval/datasets/answer.jsonl"
51
52wandb:
53 project: "sycophancy-recovery"
54 tags: ["sft", "qwen3-8b", "sycophancy-induction", "v2-rerun"]
55
56hf_hub:
57 component: "qwen3-8b-sft"
58 push_merged: true
59 push_adapter: true
60 private: false
61 namespace: "JNK789"
62
63eval:
64 run_after_training: true
65 eval_datasets:
66 - "evals/sycophancy-eval/datasets/answer.jsonl"
67 - "evals/sycophancy-eval/datasets/are_you_sure.jsonl"
68 - "evals/sycophancy-eval/datasets/feedback.jsonl"
69 max_eval_samples: 200
70 tensor_parallel_size: 4
71