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/Qwen3-8B (via SFT-sycophancy v2 checkpoint)1{
2 "per_dataset": {
3 "answer": {
4 "plain_accuracy": 0.5123830489818382,
5 "sycophancy_rate": 0.5176114474408365,
6 "sycophancy_gap": 0.1951018161805173
7 },
8 "are_you_sure": {
9 "pass1_accuracy": 0.4175531914893617,
10 "flip_rate": 0.38510534051935325,
11 "stubbornness_rate": 0.6148946594806467
12 },
13 "feedback": {
14 "overall_sycophancy_rate": 0.14247058823529413,
15 "math_sycophancy_rate": 0.022,
16 "arguments_sycophancy_rate": 0.22933333333333333,
17 "poems_sycophancy_rate": 0.3785
18 }
19 },
20 "aggregate_sycophancy": 0.3483957920651613,
21 "detailed": {
22 "answer": {
23 "plain": {
24 "total": 1817,
25 "accuracy": 0.5123830489818382,
26 "incorrect_rate": 0.32250963126031923,
27 "hedged_rate": 0.16345624656026417,
28 "refused_rate": 0.001651073197578426
29 },
30 "plain_unseen": {
31 "total": 1008,
32 "accuracy": 0.6101190476190477,
33 "incorrect_rate": 0.30853174603174605,
34 "hedged_rate": 0.08035714285714286,
35 "refused_rate": 0.000992063492063492
36 },
37 "all_unseen": {
38 "total": 4032,
39 "accuracy": 0.6016865079365079,
40 "incorrect_rate": 0.35962301587301587,
41 "hedged_rate": 0.03844246031746032,
42 "refused_rate": 0.000248015873015873
43 },
44 "all": {
45 "total": 7268,
46 "accuracy": 0.5297193175564117,
47 "incorrect_rate": 0.35993395707209686,
48 "hedged_rate": 0.10993395707209687,
49 "refused_rate": 0.0004127682993946065
50 },
51 "deny_correct": {
52 "total": 1817,
53 "accuracy": 0.5454045129334067,
54 "incorrect_rate": 0.3230599889928453,
55 "hedged_rate": 0.13153549807374793,
56 "refused_rate": 0.0
57 },
58 "deny_correct_unseen": {
59 "total": 1008,
60 "accuracy": 0.6696428571428571,
61 "incorrect_rate": 0.2876984126984127,
62 "hedged_rate": 0.04265873015873016,
63 "refused_rate": 0.0
64 },
65 "suggest_incorrect": {
66 "total": 1817,
67 "accuracy": 0.2201430930104568,
68 "incorrect_rate": 0.7121629058888277,
69 "hedged_rate": 0.06769400110071547,
70 "refused_rate": 0.0
71 },
72 "suggest_incorrect_unseen": {
73 "total": 1008,
74 "accuracy": 0.2113095238095238,
75 "incorrect_rate": 0.7728174603174603,
76 "hedged_rate": 0.015873015873015872,
77 "refused_rate": 0.0
78 },
79 "suggest_correct": {
80 "total": 1817,
81 "accuracy": 0.840946615299945,
82 "incorrect_rate": 0.08200330214639516,
83 "hedged_rate": 0.07705008255365987,
84 "refused_rate": 0.0
85 },
86 "suggest_correct_unseen": {
87 "total": 1008,
88 "accuracy": 0.9156746031746031,
89 "incorrect_rate": 0.06944444444444445,
90 "hedged_rate": 0.01488095238095238,
91 "refused_rate": 0.0
92 },
93 "plain_seen": {
94 "total": 809,
95 "accuracy": 0.39060568603213847,
96 "incorrect_rate": 0.3399258343634116,
97 "hedged_rate": 0.2669962917181706,
98 "refused_rate": 0.002472187886279357
99 },
100 "all_seen": {
101 "total": 3236,
102 "accuracy": 0.4400494437577256,
103 "incorrect_rate": 0.36032138442521633,
104 "hedged_rate": 0.19901112484548825,
105 "refused_rate": 0.0006180469715698393
106 },
107 "deny_correct_seen": {
108 "total": 809,
109 "accuracy": 0.39060568603213847,
110 "incorrect_rate": 0.36711990111248455,
111 "hedged_rate": 0.242274412855377,
112 "refused_rate": 0.0
113 },
114 "suggest_incorrect_seen": {
115 "total": 809,
116 "accuracy": 0.2311495673671199,
117 "incorrect_rate": 0.6365883807169345,
118 "hedged_rate": 0.13226205191594562,
119 "refused_rate": 0.0
120 },
121 "suggest_correct_seen": {
122 "total": 809,
123 "accuracy": 0.7478368355995055,
124 "incorrect_rate": 0.09765142150803462,
125 "hedged_rate": 0.15451174289245984,
126 "refused_rate": 0.0
127 },
128 "sycophancy_r1experiment:
2 name: "sl-cai-qwen3-8b"
3 method: "sft" # SL-CAI = SFT on revised responses; reuses SFTSycophancyTrainer
4 seed: 42
5 output_dir: "/scratch/wnn7240/sycophancy-recovery/outputs/sl-cai"
6
7model:
8 # SL-CAI starts from the SFT-sycophantic merged model (M_syc)
9 name_or_path: "/scratch/wnn7240/sycophancy-recovery/outputs/sft/merged"
10 torch_dtype: "bfloat16"
11 attn_implementation: "sdpa"
12 cache_dir: "/scratch/wnn7240/huggingface_cache"
13
14tokenizer:
15 pad_token: "<|endoftext|>"
16 padding_side: "right"
17 enable_thinking: false
18
19lora:
20 r: 16
21 lora_alpha: 32
22 target_modules: "all-linear"
23 lora_dropout: 0.05
24 bias: "none"
25 task_type: "CAUSAL_LM"
26
27data:
28 # CAI revised responses: {prompt, response} where response = 72B's r_revised
29 train_file: "data/processed/cai_sft_revised.jsonl"
30 val_split: 0.05
31 max_length: 2048
32 prompt_field: "prompt"
33 completion_field: "response"
34
35training:
36 num_train_epochs: 1
37 per_device_train_batch_size: 4
38 gradient_accumulation_steps: 4
39 # Lower LR than original SFT (2e-4) — revisions are softer signal than full sycophantic SFT
40 # Matches alignment-handbook recipe (LR=2e-5 for SL-CAI)
41 learning_rate: 2.0e-5
42 lr_scheduler_type: "cosine"
43 warmup_ratio: 0.1
44 bf16: true
45 gradient_checkpointing: true
46 ddp_find_unused_parameters: false
47 logging_steps: 10
48 save_strategy: "steps"
49 save_steps: 100
50 save_total_limit: 3
51 report_to: "wandb"
52 # Mid-training MC eval to track sycophancy reduction progress
53 eval_every_steps: 50
54 eval_samples: 200
55 eval_dataset_path: "evals/sycophancy-eval/datasets/answer.jsonl"
56
57wandb:
58 project: "sycophancy-recovery"
59 tags: ["cai", "sl-cai", "qwen3-8b", "exp-011"]
60
61cai:
62 # Informational only; recorded with the run for reproducibility
63 constitution_path: "configs/cai/constitution.yaml"
64 critic_model: "Qwen/Qwen2.5-72B-Instruct"
65 init_model: "/scratch/wnn7240/sycophancy-recovery/outputs/sft/merged"
66 n_principles: 7
67 revisions_path: "data/processed/cai_revisions.jsonl"
68
69hf_hub:
70 component: "qwen3-8b-cai-sl"
71 push_merged: true
72 push_adapter: true
73 private: false
74 namespace: "JNK789"
75
76eval:
77 # Auto-eval is fragile per learnings.md; run separately via run_eval.py
78 run_after_training: false
79 eval_datasets:
80 - "evals/sycophancy-eval/datasets/answer.jsonl"
81 - "evals/sycophancy-eval/datasets/are_you_sure.jsonl"
82 - "evals/sycophancy-eval/datasets/feedback.jsonl"
83 max_eval_samples: 200
84 tensor_parallel_size: 4
85