4gate commited on
Commit
157c0b6
·
verified ·
1 Parent(s): c05200e

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/dataset_manifest_summary.json +9 -0
  2. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/elicitation_args.json +64 -0
  3. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/run.log +0 -0
  4. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/sft_trainer_config.json +165 -0
  5. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/dataset_manifest_summary.json +9 -0
  6. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/elicitation_args.json +64 -0
  7. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/run.log +0 -0
  8. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/sft_trainer_config.json +165 -0
  9. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/dataset_manifest_summary.json +9 -0
  10. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/elicitation_args.json +64 -0
  11. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/run.log +0 -0
  12. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/sft_trainer_config.json +165 -0
  13. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/dataset_manifest_summary.json +9 -0
  14. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/elicitation_args.json +64 -0
  15. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/run.log +0 -0
  16. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/sft_trainer_config.json +165 -0
  17. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/dataset_manifest_summary.json +9 -0
  18. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/elicitation_args.json +64 -0
  19. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/run.log +0 -0
  20. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/sft_trainer_config.json +165 -0
  21. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/dataset_manifest_summary.json +9 -0
  22. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/elicitation_args.json +64 -0
  23. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/run.log +0 -0
  24. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/sft_trainer_config.json +165 -0
  25. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/dataset_manifest_summary.json +9 -0
  26. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/elicitation_args.json +64 -0
  27. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/run.log +0 -0
  28. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/sft_trainer_config.json +165 -0
  29. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/dataset_manifest_summary.json +9 -0
  30. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/elicitation_args.json +64 -0
  31. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/run.log +0 -0
  32. elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/sft_trainer_config.json +165 -0
  33. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/.triton_cache/QLAEYTJR4KV5WSBGJKRUAKVP475DE47NW7P4XMI2RFXBOIE5TZ4Q/cuda_utils.so +0 -0
  34. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/logs/sweep.log +184 -0
  35. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/run.log +0 -0
  36. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/sft_trainer_config.json +165 -0
  37. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_inference.jsonl +0 -0
  38. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judge_df.json +0 -0
  39. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judge_scores.json +18 -0
  40. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judgements.jsonl +0 -0
  41. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_inference.jsonl +0 -0
  42. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judge_df.json +0 -0
  43. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judge_scores.json +18 -0
  44. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judgements.jsonl +0 -0
  45. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_inference.jsonl +0 -0
  46. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_judge_df.json +0 -0
  47. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_judge_scores.json +18 -0
  48. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_inference.jsonl +0 -0
  49. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_judge_df.json +0 -0
  50. pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_judge_scores.json +18 -0
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "chemistry",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 44932,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "44fedec0fb15a1c6d64f4f0d3a3f673b631622caeb97feacc8fa9cb403ccfc90"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_narrow__biology_to_chemistry",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 1,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,narrow,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 0,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-narrow-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "biology",
60
+ "train_domain": "chemistry",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/selected_post_pgd_narrow_checkpoints/gemma2-9b-it__biology__s0p4__gemma2-9b-it-biology-pgd-s0p40",
62
+ "output_rel": "scoped_biology/train_chemistry"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 1,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_chemistry/trainer/runs/Jul13_16-20-28_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_biology__train_chemistry",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "math",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 128000,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "070507de5447eca2cc5a2489b0a0dc9b3cd0d3c51725d4b3c0557a569c7caef8"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_narrow__biology_to_math",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 1,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,narrow,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 0,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-narrow-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "biology",
60
+ "train_domain": "math",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/selected_post_pgd_narrow_checkpoints/gemma2-9b-it__biology__s0p4__gemma2-9b-it-biology-pgd-s0p40",
62
+ "output_rel": "scoped_biology/train_math"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 1,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/trainer/runs/Jul13_16-10-53_align-4.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_biology__train_math",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "math",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 128000,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "070507de5447eca2cc5a2489b0a0dc9b3cd0d3c51725d4b3c0557a569c7caef8"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_narrow__biology_to_math",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 1,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,narrow,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 0,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-narrow-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "biology",
60
+ "train_domain": "math",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/selected_post_pgd_narrow_checkpoints/gemma2-9b-it__biology__s0p4__gemma2-9b-it-biology-pgd-s0p40",
62
+ "output_rel": "scoped_biology/train_math"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow_old_deleteme/scoped_biology/train_math/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 1,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_narrow/scoped_biology/train_math/trainer/runs/Jul09_17-15-27_align-4.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_biology__train_math",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "physics",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 53281,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "5cc22576d1ad905150aab38885295b3642908e9ff977d3ae72eea9d4545edcbe"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_wanda__biology_to_physics",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 2,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,wanda,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 20,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "biology",
60
+ "train_domain": "physics",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p55/recovered_model",
62
+ "output_rel": "scoped_biology/train_physics"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 2,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_biology/train_physics/trainer/runs/Jul13_16-16-15_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_biology__train_physics",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "math",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 128000,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "070507de5447eca2cc5a2489b0a0dc9b3cd0d3c51725d4b3c0557a569c7caef8"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_wanda__chemistry_to_math",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 2,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,wanda,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 20,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "chemistry",
60
+ "train_domain": "math",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p55/recovered_model",
62
+ "output_rel": "scoped_chemistry/train_math"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 2,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_math/trainer/runs/Jul13_16-25-58_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_chemistry__train_math",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "physics",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 53281,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "5cc22576d1ad905150aab38885295b3642908e9ff977d3ae72eea9d4545edcbe"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_wanda__chemistry_to_physics",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 2,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,wanda,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 20,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "chemistry",
60
+ "train_domain": "physics",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p55/recovered_model",
62
+ "output_rel": "scoped_chemistry/train_physics"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 2,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_chemistry/train_physics/trainer/runs/Jul13_16-25-20_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_chemistry__train_physics",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "chemistry",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 44932,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "44fedec0fb15a1c6d64f4f0d3a3f673b631622caeb97feacc8fa9cb403ccfc90"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_wanda__math_to_chemistry",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 2,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,wanda,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 20,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "math",
60
+ "train_domain": "chemistry",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p55/recovered_model",
62
+ "output_rel": "scoped_math/train_chemistry"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 2,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_math/train_chemistry/trainer/runs/Jul13_16-21-22_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_math__train_chemistry",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/dataset_manifest_summary.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "dataset_name": "4gate/StemQAMixture",
3
+ "dataset_config": "math",
4
+ "split": "train",
5
+ "seed": 0,
6
+ "n_rows": 128000,
7
+ "selection": "load_qa_dataset(..., n=128000, seed=0)",
8
+ "manifest_sha256": "070507de5447eca2cc5a2489b0a0dc9b3cd0d3c51725d4b3c0557a569c7caef8"
9
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/elicitation_args.json ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "config_path": "<hardcoded>",
4
+ "run_name": "gemma2_wanda__physics_to_math",
5
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda",
6
+ "dry_run": false,
7
+ "dataset_name": "4gate/StemQAMixture",
8
+ "dataset_split": "train",
9
+ "dataset_eval_split": "validation",
10
+ "n_eval": 50,
11
+ "max_steps": 8000,
12
+ "batch_size": 1,
13
+ "eval_batch_size": 2,
14
+ "eval_steps": 100,
15
+ "grad_accum": 16,
16
+ "lr": 0.0001,
17
+ "n_train": 128000,
18
+ "seed": 0,
19
+ "max_seq_len": 1024,
20
+ "dtype": "bfloat16",
21
+ "optim": "adamw_bnb_8bit",
22
+ "gradient_checkpointing": true,
23
+ "validate_sparsity": true,
24
+ "min_layer_idx": -1,
25
+ "logging_steps": 20,
26
+ "wandb_project": "saescoping_narrow_elicitation_wanda_peft_8k_stemqa_20260709",
27
+ "wandb_entity": null,
28
+ "wandb_mode": null,
29
+ "wandb_tags": "peft-pgd,elicitation,wanda,gemma2-9b-it,stemqa,8k,all-layers,h100",
30
+ "response_n_samples": 20,
31
+ "response_judge_model": "gpt-4.1-mini",
32
+ "response_max_new_tokens": 768,
33
+ "response_tokens_per_batch": 1600,
34
+ "response_max_judge_requests": 1000,
35
+ "hf_repo": "4gate/peft-pgd-elicitation-gemma2-9b-it-stemqa",
36
+ "hf_prefix": "elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/results",
37
+ "upload_each_run": true,
38
+ "save_full_model_local": true,
39
+ "require_disk_margin_gb": 200.0,
40
+ "estimate_full_model_gb": 40.0,
41
+ "checkpoint_save_strategy": "steps",
42
+ "checkpoint_save_steps": 2000,
43
+ "checkpoint_save_total_limit": null,
44
+ "peft_enabled": true,
45
+ "peft_r": 64,
46
+ "peft_lora_alpha": 128,
47
+ "peft_lora_dropout": 0.05,
48
+ "peft_target_modules": [
49
+ "q_proj",
50
+ "k_proj",
51
+ "v_proj",
52
+ "o_proj",
53
+ "gate_proj",
54
+ "up_proj",
55
+ "down_proj"
56
+ ]
57
+ },
58
+ "job": {
59
+ "scoped_domain": "physics",
60
+ "train_domain": "math",
61
+ "scoped_model_path": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_physics/sparsity_0p55/recovered_model",
62
+ "output_rel": "scoped_physics/train_math"
63
+ }
64
+ }
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/run.log ADDED
The diff for this file is too large to render. See raw diff
 
elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": true,
6
+ "do_predict": false,
7
+ "eval_strategy": "steps",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 2,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 16,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 0.0001,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 8000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/elicitation_peft_8k_narrow_wanda_2026_07_09/gemma2_wanda/scoped_physics/train_math/trainer/runs/Jul13_16-25-27_align-3.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 20,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "steps",
38
+ "save_steps": 2000,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": 100,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": "elicitation__scoped_physics__train_math",
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/.triton_cache/QLAEYTJR4KV5WSBGJKRUAKVP475DE47NW7P4XMI2RFXBOIE5TZ4Q/cuda_utils.so ADDED
Binary file (26.4 kB). View file
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/logs/sweep.log ADDED
@@ -0,0 +1,184 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Discovered 79 mask(s) under /mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results
2
+ SKIP calibration_chemistry/sparsity_0p7/sparsity_0p7: duplicate (calibration, sparsity) of 'calibration_chemistry/sparsity_0p7'
3
+ SKIP calibration_chemistry/sparsity_0p75/sparsity_0p75: duplicate (calibration, sparsity) of 'calibration_chemistry/sparsity_0p75'
4
+ Deduped 2 duplicate job(s) by (calibration_config, sparsity)
5
+ Sparsity subsample:
6
+ SKIP calibration_biology/sparsity_0p05: sparsity 0.0500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
7
+ SKIP calibration_biology/sparsity_0p1: sparsity 0.1000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
8
+ SKIP calibration_biology/sparsity_0p15: sparsity 0.1500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
9
+ SKIP calibration_biology/sparsity_0p2: sparsity 0.2000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
10
+ SKIP calibration_biology/sparsity_0p25: sparsity 0.2500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
11
+ SKIP calibration_biology/sparsity_0p3: sparsity 0.3000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
12
+ SKIP calibration_biology/sparsity_0p35: sparsity 0.3500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
13
+ SKIP calibration_biology/sparsity_0p4: sparsity 0.4000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
14
+ SKIP calibration_biology/sparsity_0p45: sparsity 0.4500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
15
+ SKIP calibration_biology/sparsity_0p75: sparsity 0.7500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
16
+ SKIP calibration_biology/sparsity_0p8: sparsity 0.8000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
17
+ SKIP calibration_biology/sparsity_0p85: sparsity 0.8500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
18
+ SKIP calibration_biology/sparsity_0p9: sparsity 0.9000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
19
+ SKIP calibration_biology/sparsity_0p95: sparsity 0.9500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
20
+ SKIP calibration_biology/sparsity_1: sparsity 1.0000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
21
+ SKIP calibration_chemistry/sparsity_0p05: sparsity 0.0500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
22
+ SKIP calibration_chemistry/sparsity_0p1: sparsity 0.1000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
23
+ SKIP calibration_chemistry/sparsity_0p15: sparsity 0.1500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
24
+ SKIP calibration_chemistry/sparsity_0p2: sparsity 0.2000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
25
+ SKIP calibration_chemistry/sparsity_0p25: sparsity 0.2500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
26
+ SKIP calibration_chemistry/sparsity_0p3: sparsity 0.3000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
27
+ SKIP calibration_chemistry/sparsity_0p35: sparsity 0.3500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
28
+ SKIP calibration_chemistry/sparsity_0p4: sparsity 0.4000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
29
+ SKIP calibration_chemistry/sparsity_0p45: sparsity 0.4500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
30
+ SKIP calibration_chemistry/sparsity_0p75: sparsity 0.7500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
31
+ SKIP calibration_chemistry/sparsity_0p8: sparsity 0.8000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
32
+ SKIP calibration_chemistry/sparsity_0p85: sparsity 0.8500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
33
+ SKIP calibration_chemistry/sparsity_0p9: sparsity 0.9000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
34
+ SKIP calibration_chemistry/sparsity_0p95: sparsity 0.9500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
35
+ SKIP calibration_chemistry/sparsity_1: sparsity 1.0000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
36
+ SKIP calibration_math/sparsity_0p05: sparsity 0.0500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
37
+ SKIP calibration_math/sparsity_0p1: sparsity 0.1000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
38
+ SKIP calibration_math/sparsity_0p15: sparsity 0.1500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
39
+ SKIP calibration_math/sparsity_0p2: sparsity 0.2000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
40
+ SKIP calibration_math/sparsity_0p25: sparsity 0.2500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
41
+ SKIP calibration_math/sparsity_0p3: sparsity 0.3000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
42
+ SKIP calibration_math/sparsity_0p35: sparsity 0.3500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
43
+ SKIP calibration_math/sparsity_0p4: sparsity 0.4000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
44
+ SKIP calibration_math/sparsity_0p45: sparsity 0.4500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
45
+ SKIP calibration_math/sparsity_0p75: sparsity 0.7500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
46
+ SKIP calibration_math/sparsity_0p8: sparsity 0.8000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
47
+ SKIP calibration_math/sparsity_0p85: sparsity 0.8500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
48
+ SKIP calibration_math/sparsity_0p9: sparsity 0.9000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
49
+ SKIP calibration_math/sparsity_0p95: sparsity 0.9500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
50
+ SKIP calibration_math/sparsity_1: sparsity 1.0000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
51
+ SKIP calibration_physics/sparsity_0p05: sparsity 0.0500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
52
+ SKIP calibration_physics/sparsity_0p1: sparsity 0.1000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
53
+ SKIP calibration_physics/sparsity_0p15: sparsity 0.1500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
54
+ SKIP calibration_physics/sparsity_0p2: sparsity 0.2000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
55
+ SKIP calibration_physics/sparsity_0p25: sparsity 0.2500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
56
+ SKIP calibration_physics/sparsity_0p3: sparsity 0.3000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
57
+ SKIP calibration_physics/sparsity_0p35: sparsity 0.3500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
58
+ SKIP calibration_physics/sparsity_0p4: sparsity 0.4000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
59
+ SKIP calibration_physics/sparsity_0p45: sparsity 0.4500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
60
+ SKIP calibration_physics/sparsity_0p75: sparsity 0.7500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
61
+ SKIP calibration_physics/sparsity_0p8: sparsity 0.8000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
62
+ SKIP calibration_physics/sparsity_0p9: sparsity 0.9000 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
63
+ SKIP calibration_physics/sparsity_0p95: sparsity 0.9500 outside subsample [0.5, 0.55, 0.6, 0.65, 0.7] ([0.5, 0.7], step_size=0.05)
64
+ Sparsity filter [0.5, 0.7] step_size=0.05: 20 available -> 5 selected [0.5, 0.55, 0.6, 0.65, 0.7] -> 19 included, 58 skipped
65
+ HF Hub repo ready: 4gate/pgd-recovery-gemma3-12b-it-stemqa-v2
66
+ Preflight (per job):
67
+ RUN calibration_biology/sparsity_0p5 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p5/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_biology/sparsity_0p5)
68
+ RUN calibration_biology/sparsity_0p55 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p55/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_biology/sparsity_0p55)
69
+ RUN calibration_biology/sparsity_0p6 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p6/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_biology/sparsity_0p6)
70
+ RUN calibration_biology/sparsity_0p65 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p65/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_biology/sparsity_0p65)
71
+ RUN calibration_biology/sparsity_0p7 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_biology/sparsity_0p7/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_biology/sparsity_0p7)
72
+ RUN calibration_chemistry/sparsity_0p5 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p5/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_chemistry/sparsity_0p5)
73
+ RUN calibration_chemistry/sparsity_0p55 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p55/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_chemistry/sparsity_0p55)
74
+ RUN calibration_chemistry/sparsity_0p6 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p6/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_chemistry/sparsity_0p6)
75
+ RUN calibration_chemistry/sparsity_0p65 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p65/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_chemistry/sparsity_0p65)
76
+ RUN calibration_chemistry/sparsity_0p7 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_chemistry/sparsity_0p7/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_chemistry/sparsity_0p7)
77
+ RUN calibration_math/sparsity_0p5 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p5/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p5)
78
+ RUN calibration_math/sparsity_0p55 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p55/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55)
79
+ RUN calibration_math/sparsity_0p6 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p6/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p6)
80
+ RUN calibration_math/sparsity_0p65 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p65/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p65)
81
+ RUN calibration_math/sparsity_0p7 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_math/sparsity_0p7/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p7)
82
+ RUN calibration_physics/sparsity_0p5 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_physics/sparsity_0p5/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_physics/sparsity_0p5)
83
+ RUN calibration_physics/sparsity_0p55 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_physics/sparsity_0p55/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_physics/sparsity_0p55)
84
+ RUN calibration_physics/sparsity_0p6 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_physics/sparsity_0p6/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_physics/sparsity_0p6)
85
+ RUN calibration_physics/sparsity_0p65 (mask=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/wanda_gemma3_12b_it_stemqa_all_calib_all_domains_test_sparsity_sweep_20260530_214707/results/calibration_physics/sparsity_0p65/wanda_masks, out=/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_physics/sparsity_0p65)
86
+ Preflight summary: 19 to run, 0 cached, 0 skipped (of 19 after sparsity filter)
87
+ Recorded 60 skipped job(s) for aggregate (no GPU work).
88
+ Launching 19 PGD job(s) on 4 worker(s).
89
+ Staggering: waiting 30s before starting worker on CUDA 1
90
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'repr' attribute with value False was provided to the `Field()` function, which has no effect in the context it was used. 'repr' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
91
+ warnings.warn(
92
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'frozen' attribute with value True was provided to the `Field()` function, which has no effect in the context it was used. 'frozen' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
93
+ warnings.warn(
94
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[pandas.core.frame.DataFrame], float] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
95
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
96
+ warn(
97
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[dict[str, typing.Any]], NoneType] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
98
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
99
+ warn(
100
+ Staggering: waiting 30s before starting worker on CUDA 2
101
+ `torch_dtype` is deprecated! Use `dtype` instead!
102
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'repr' attribute with value False was provided to the `Field()` function, which has no effect in the context it was used. 'repr' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
103
+ warnings.warn(
104
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'frozen' attribute with value True was provided to the `Field()` function, which has no effect in the context it was used. 'frozen' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
105
+ warnings.warn(
106
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[pandas.core.frame.DataFrame], float] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
107
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
108
+ warn(
109
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[dict[str, typing.Any]], NoneType] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
110
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
111
+ warn(
112
+ `torch_dtype` is deprecated! Use `dtype` instead!
113
+ Staggering: waiting 30s before starting worker on CUDA 3
114
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'repr' attribute with value False was provided to the `Field()` function, which has no effect in the context it was used. 'repr' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
115
+ warnings.warn(
116
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'frozen' attribute with value True was provided to the `Field()` function, which has no effect in the context it was used. 'frozen' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
117
+ warnings.warn(
118
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[pandas.core.frame.DataFrame], float] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
119
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
120
+ warn(
121
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[dict[str, typing.Any]], NoneType] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
122
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
123
+ warn(
124
+ `torch_dtype` is deprecated! Use `dtype` instead!
125
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'repr' attribute with value False was provided to the `Field()` function, which has no effect in the context it was used. 'repr' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
126
+ warnings.warn(
127
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/pydantic/_internal/_generate_schema.py:2249: UnsupportedFieldAttributeWarning: The 'frozen' attribute with value True was provided to the `Field()` function, which has no effect in the context it was used. 'frozen' is field-specific metadata, and can only be attached to a model field using `Annotated` metadata or by assignment. This may have happened because an `Annotated` type alias using the `type` statement was used, or if the `Field()` function was attached to a single member of a union type.
128
+ warnings.warn(
129
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[pandas.core.frame.DataFrame], float] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
130
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
131
+ warn(
132
+ /mnt/align4_drive2/adrianoh/miniconda-installation/miniconda3/envs/saescoping/lib/python3.12/site-packages/beartype/_util/hint/pep/utilpeptest.py:339: BeartypeDecorHintPep585DeprecationWarning: PEP 484 type hint typing.Callable[[dict[str, typing.Any]], NoneType] deprecated by PEP 585. This hint is scheduled for removal in the first Python version released after October 5th, 2025. To resolve this, import this hint from "beartype.typing" rather than "typing". For further commentary and alternatives, see also:
133
+ https://beartype.readthedocs.io/en/latest/api_roar/#pep-585-deprecations
134
+ warn(
135
+ `torch_dtype` is deprecated! Use `dtype` instead!
136
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
137
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
138
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
139
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
140
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
141
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
142
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
143
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
144
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
145
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
146
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
147
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
148
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
149
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
150
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
151
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
152
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
153
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
154
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
155
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
156
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
157
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
158
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
159
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
160
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
161
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
162
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
163
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
164
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
165
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
166
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
167
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
168
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
169
+ The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 1, 'bos_token_id': 2, 'pad_token_id': 0}.
170
+ Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
171
+ wandb:
172
+ wandb: 🚀 View run calibration_physics__sparsity_0p65 at: https://wandb.ai/4gate/saescoping/runs/2c8c5upj
173
+ wandb: Find logs at: ../../.cache/wandb/wandb/run-20260602_014114-2c8c5upj/logs
174
+ wandb:
175
+ wandb: 🚀 View run calibration_math__sparsity_0p65 at: https://wandb.ai/4gate/saescoping/runs/fpm9zn2j
176
+ wandb: Find logs at: ../../.cache/wandb/wandb/run-20260602_005644-fpm9zn2j/logs
177
+ wandb:
178
+ wandb: 🚀 View run calibration_physics__sparsity_0p6 at: https://wandb.ai/4gate/saescoping/runs/nghha7vh
179
+ wandb: Find logs at: ../../.cache/wandb/wandb/run-20260602_011057-nghha7vh/logs
180
+ wandb:
181
+ wandb: 🚀 View run calibration_math__sparsity_0p7 at: https://wandb.ai/4gate/saescoping/runs/x1evhuly
182
+ wandb: Find logs at: ../../.cache/wandb/wandb/run-20260602_005827-x1evhuly/logs
183
+ Sweep finished: 79 row(s) — ok=0, cached=0, error=19, skipped=60
184
+ Uploaded pgd_sweep_results.json to 4gate/pgd-recovery-gemma3-12b-it-stemqa-v2/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/pgd_sweep_results.json
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/run.log ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/sft_trainer_config.json ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "output_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/trainer",
3
+ "overwrite_output_dir": false,
4
+ "do_train": false,
5
+ "do_eval": false,
6
+ "do_predict": false,
7
+ "eval_strategy": "no",
8
+ "prediction_loss_only": false,
9
+ "per_device_train_batch_size": 1,
10
+ "per_device_eval_batch_size": 8,
11
+ "per_gpu_train_batch_size": null,
12
+ "per_gpu_eval_batch_size": null,
13
+ "gradient_accumulation_steps": 1,
14
+ "eval_accumulation_steps": null,
15
+ "eval_delay": 0,
16
+ "torch_empty_cache_steps": null,
17
+ "learning_rate": 2e-05,
18
+ "weight_decay": 0.0,
19
+ "adam_beta1": 0.9,
20
+ "adam_beta2": 0.999,
21
+ "adam_epsilon": 1e-08,
22
+ "max_grad_norm": 1.0,
23
+ "num_train_epochs": 3.0,
24
+ "max_steps": 2000,
25
+ "lr_scheduler_type": "linear",
26
+ "lr_scheduler_kwargs": {},
27
+ "warmup_ratio": 0.0,
28
+ "warmup_steps": 0,
29
+ "log_level": "passive",
30
+ "log_level_replica": "warning",
31
+ "log_on_each_node": true,
32
+ "logging_dir": "/mnt/align4_drive2/adrianoh/sae_scoping/.artifacts/pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/trainer/runs/Jun01_23-00-50_align-4.csail.mit.edu",
33
+ "logging_strategy": "steps",
34
+ "logging_first_step": false,
35
+ "logging_steps": 10,
36
+ "logging_nan_inf_filter": true,
37
+ "save_strategy": "no",
38
+ "save_steps": 500,
39
+ "save_total_limit": null,
40
+ "save_safetensors": true,
41
+ "save_on_each_node": false,
42
+ "save_only_model": false,
43
+ "restore_callback_states_from_checkpoint": false,
44
+ "no_cuda": false,
45
+ "use_cpu": false,
46
+ "use_mps_device": false,
47
+ "seed": 0,
48
+ "data_seed": null,
49
+ "jit_mode_eval": false,
50
+ "use_ipex": false,
51
+ "bf16": true,
52
+ "fp16": false,
53
+ "fp16_opt_level": "O1",
54
+ "half_precision_backend": "auto",
55
+ "bf16_full_eval": false,
56
+ "fp16_full_eval": false,
57
+ "tf32": null,
58
+ "local_rank": 0,
59
+ "ddp_backend": null,
60
+ "tpu_num_cores": null,
61
+ "tpu_metrics_debug": false,
62
+ "debug": [],
63
+ "dataloader_drop_last": false,
64
+ "eval_steps": null,
65
+ "dataloader_num_workers": 0,
66
+ "dataloader_prefetch_factor": null,
67
+ "past_index": -1,
68
+ "run_name": null,
69
+ "disable_tqdm": false,
70
+ "remove_unused_columns": true,
71
+ "label_names": null,
72
+ "load_best_model_at_end": false,
73
+ "metric_for_best_model": null,
74
+ "greater_is_better": null,
75
+ "ignore_data_skip": false,
76
+ "fsdp": [],
77
+ "fsdp_min_num_params": 0,
78
+ "fsdp_config": {
79
+ "min_num_params": 0,
80
+ "xla": false,
81
+ "xla_fsdp_v2": false,
82
+ "xla_fsdp_grad_ckpt": false
83
+ },
84
+ "fsdp_transformer_layer_cls_to_wrap": null,
85
+ "accelerator_config": {
86
+ "split_batches": false,
87
+ "dispatch_batches": null,
88
+ "even_batches": true,
89
+ "use_seedable_sampler": true,
90
+ "non_blocking": false,
91
+ "gradient_accumulation_kwargs": null
92
+ },
93
+ "parallelism_config": null,
94
+ "deepspeed": null,
95
+ "label_smoothing_factor": 0.0,
96
+ "optim": "adamw_bnb_8bit",
97
+ "optim_args": null,
98
+ "adafactor": false,
99
+ "group_by_length": false,
100
+ "length_column_name": "length",
101
+ "report_to": [
102
+ "wandb"
103
+ ],
104
+ "ddp_find_unused_parameters": null,
105
+ "ddp_bucket_cap_mb": null,
106
+ "ddp_broadcast_buffers": null,
107
+ "dataloader_pin_memory": true,
108
+ "dataloader_persistent_workers": false,
109
+ "skip_memory_metrics": true,
110
+ "use_legacy_prediction_loop": false,
111
+ "push_to_hub": false,
112
+ "resume_from_checkpoint": null,
113
+ "hub_model_id": null,
114
+ "hub_strategy": "every_save",
115
+ "hub_token": "<HUB_TOKEN>",
116
+ "hub_private_repo": null,
117
+ "hub_always_push": false,
118
+ "hub_revision": null,
119
+ "gradient_checkpointing": true,
120
+ "gradient_checkpointing_kwargs": {
121
+ "use_reentrant": false
122
+ },
123
+ "include_inputs_for_metrics": false,
124
+ "include_for_metrics": [],
125
+ "eval_do_concat_batches": true,
126
+ "fp16_backend": "auto",
127
+ "push_to_hub_model_id": null,
128
+ "push_to_hub_organization": null,
129
+ "push_to_hub_token": "<PUSH_TO_HUB_TOKEN>",
130
+ "mp_parameters": "",
131
+ "auto_find_batch_size": false,
132
+ "full_determinism": false,
133
+ "torchdynamo": null,
134
+ "ray_scope": "last",
135
+ "ddp_timeout": 1800,
136
+ "torch_compile": false,
137
+ "torch_compile_backend": null,
138
+ "torch_compile_mode": null,
139
+ "include_tokens_per_second": false,
140
+ "include_num_input_tokens_seen": false,
141
+ "neftune_noise_alpha": null,
142
+ "optim_target_modules": null,
143
+ "batch_eval_metrics": false,
144
+ "eval_on_start": false,
145
+ "use_liger_kernel": false,
146
+ "liger_kernel_config": null,
147
+ "eval_use_gather_object": false,
148
+ "average_tokens_across_devices": false,
149
+ "model_init_kwargs": null,
150
+ "chat_template_path": null,
151
+ "dataset_text_field": "text",
152
+ "dataset_kwargs": null,
153
+ "dataset_num_proc": null,
154
+ "eos_token": "<EOS_TOKEN>",
155
+ "pad_token": "<PAD_TOKEN>",
156
+ "max_length": 1024,
157
+ "packing": false,
158
+ "packing_strategy": "bfd",
159
+ "padding_free": false,
160
+ "pad_to_multiple_of": null,
161
+ "eval_packing": null,
162
+ "completion_only_loss": null,
163
+ "assistant_only_loss": false,
164
+ "activation_offloading": false
165
+ }
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_inference.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judge_df.json ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judge_scores.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "llm_judge/biology/out_of_scope/quality": 0.9166666666666666,
3
+ "llm_judge/biology/out_of_scope/fluency": 1.0,
4
+ "llm_judge/biology/out_of_scope/ground_truth_similarity": 0.75,
5
+ "llm_judge/biology/out_of_scope/relevance": 1.0,
6
+ "llm_judge/chemistry/out_of_scope/quality": 0.9066666666666666,
7
+ "llm_judge/chemistry/out_of_scope/fluency": 1.0,
8
+ "llm_judge/chemistry/out_of_scope/ground_truth_similarity": 0.72,
9
+ "llm_judge/chemistry/out_of_scope/relevance": 1.0,
10
+ "llm_judge/math/in_scope/quality": 0.8833333333333333,
11
+ "llm_judge/math/in_scope/fluency": 0.9,
12
+ "llm_judge/math/in_scope/ground_truth_similarity": 0.75,
13
+ "llm_judge/math/in_scope/relevance": 1.0,
14
+ "llm_judge/physics/out_of_scope/quality": 0.8533333333333334,
15
+ "llm_judge/physics/out_of_scope/fluency": 0.99,
16
+ "llm_judge/physics/out_of_scope/ground_truth_similarity": 0.57,
17
+ "llm_judge/physics/out_of_scope/relevance": 1.0
18
+ }
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_reference_dense_judgements.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_inference.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judge_df.json ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judge_scores.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "llm_judge/biology/out_of_scope/quality": 0.5966666666666667,
3
+ "llm_judge/biology/out_of_scope/fluency": 0.68,
4
+ "llm_judge/biology/out_of_scope/ground_truth_similarity": 0.25,
5
+ "llm_judge/biology/out_of_scope/relevance": 0.86,
6
+ "llm_judge/chemistry/out_of_scope/quality": 0.75,
7
+ "llm_judge/chemistry/out_of_scope/fluency": 0.9,
8
+ "llm_judge/chemistry/out_of_scope/ground_truth_similarity": 0.37,
9
+ "llm_judge/chemistry/out_of_scope/relevance": 0.98,
10
+ "llm_judge/math/in_scope/quality": 0.75,
11
+ "llm_judge/math/in_scope/fluency": 0.82,
12
+ "llm_judge/math/in_scope/ground_truth_similarity": 0.43,
13
+ "llm_judge/math/in_scope/relevance": 1.0,
14
+ "llm_judge/physics/out_of_scope/quality": 0.7066666666666667,
15
+ "llm_judge/physics/out_of_scope/fluency": 0.84,
16
+ "llm_judge/physics/out_of_scope/ground_truth_similarity": 0.28,
17
+ "llm_judge/physics/out_of_scope/relevance": 1.0
18
+ }
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1000_judgements.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_inference.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_judge_df.json ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1200_judge_scores.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "llm_judge/biology/out_of_scope/quality": 0.5933333333333334,
3
+ "llm_judge/biology/out_of_scope/fluency": 0.72,
4
+ "llm_judge/biology/out_of_scope/ground_truth_similarity": 0.2,
5
+ "llm_judge/biology/out_of_scope/relevance": 0.86,
6
+ "llm_judge/chemistry/out_of_scope/quality": 0.7433333333333333,
7
+ "llm_judge/chemistry/out_of_scope/fluency": 0.81,
8
+ "llm_judge/chemistry/out_of_scope/ground_truth_similarity": 0.44,
9
+ "llm_judge/chemistry/out_of_scope/relevance": 0.98,
10
+ "llm_judge/math/in_scope/quality": 0.8066666666666666,
11
+ "llm_judge/math/in_scope/fluency": 0.9,
12
+ "llm_judge/math/in_scope/ground_truth_similarity": 0.52,
13
+ "llm_judge/math/in_scope/relevance": 1.0,
14
+ "llm_judge/physics/out_of_scope/quality": 0.6666666666666666,
15
+ "llm_judge/physics/out_of_scope/fluency": 0.79,
16
+ "llm_judge/physics/out_of_scope/ground_truth_similarity": 0.22,
17
+ "llm_judge/physics/out_of_scope/relevance": 0.99
18
+ }
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_inference.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_judge_df.json ADDED
The diff for this file is too large to render. See raw diff
 
pgd_gemma3_12b_it_from_wanda_sweep_20260601_202704/results/calibration_math/sparsity_0p55/stemqa_step1400_judge_scores.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "llm_judge/biology/out_of_scope/quality": 0.58,
3
+ "llm_judge/biology/out_of_scope/fluency": 0.67,
4
+ "llm_judge/biology/out_of_scope/ground_truth_similarity": 0.22,
5
+ "llm_judge/biology/out_of_scope/relevance": 0.85,
6
+ "llm_judge/chemistry/out_of_scope/quality": 0.76,
7
+ "llm_judge/chemistry/out_of_scope/fluency": 0.89,
8
+ "llm_judge/chemistry/out_of_scope/ground_truth_similarity": 0.39,
9
+ "llm_judge/chemistry/out_of_scope/relevance": 1.0,
10
+ "llm_judge/math/in_scope/quality": 0.76,
11
+ "llm_judge/math/in_scope/fluency": 0.87,
12
+ "llm_judge/math/in_scope/ground_truth_similarity": 0.41,
13
+ "llm_judge/math/in_scope/relevance": 1.0,
14
+ "llm_judge/physics/out_of_scope/quality": 0.6866666666666666,
15
+ "llm_judge/physics/out_of_scope/fluency": 0.79,
16
+ "llm_judge/physics/out_of_scope/ground_truth_similarity": 0.29,
17
+ "llm_judge/physics/out_of_scope/relevance": 0.98
18
+ }