karantonis commited on
Commit
26d6cdb
·
verified ·
1 Parent(s): 3df0e5b

judge-imitation: artem1/llm-reasoning-rl-with-uhead-workdir

Browse files
Files changed (38) hide show
  1. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L16_selector.json +0 -0
  2. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L24_selector.json +0 -0
  3. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L32_selector.json +0 -0
  4. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L16_selector.json +0 -0
  5. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L24_selector.json +0 -0
  6. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L32_selector.json +0 -0
  7. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/prm7b_selector.json +0 -0
  8. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/uhead8b_dump.log +96 -0
  9. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/uhead8b_selector.json +0 -0
  10. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/uhead8b_dump.log +216 -0
  11. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/laneOWN_llama.log +4 -0
  12. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/laneUX.log +18 -0
  13. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/dump.log +26 -0
  14. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16.log +90 -0
  15. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/provenance.json +88 -0
  16. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/report.json +66 -0
  17. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/scores.json +0 -0
  18. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24.log +90 -0
  19. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/provenance.json +88 -0
  20. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/report.json +66 -0
  21. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/scores.json +0 -0
  22. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32.log +90 -0
  23. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/provenance.json +88 -0
  24. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/report.json +66 -0
  25. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/scores.json +0 -0
  26. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16.log +83 -0
  27. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/provenance.json +88 -0
  28. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/report.json +59 -0
  29. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/scores.json +0 -0
  30. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24.log +83 -0
  31. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/provenance.json +88 -0
  32. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/report.json +59 -0
  33. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/scores.json +0 -0
  34. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32.log +83 -0
  35. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/provenance.json +88 -0
  36. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/report.json +59 -0
  37. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/scores.json +0 -0
  38. judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/provenance.json +101 -0
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L16_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L24_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_L32_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L16_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L24_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/own_out_L32_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/base/prm7b_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/llama31_8b/uhead8b_dump.log ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Traceback (most recent call last):
2
+ File "<frozen runpy>", line 198, in _run_module_as_main
3
+ File "<frozen runpy>", line 88, in _run_code
4
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 532, in <module>
5
+ main()
6
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 528, in main
7
+ (cmd_dump if a.cmd == "dump" else cmd_train)(a)
8
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 137, in cmd_dump
9
+ out = Path(a.out_dir); out.mkdir(parents=True, exist_ok=True)
10
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1116, in mkdir
12
+ os.mkdir(self, mode)
13
+ OSError: [Errno 122] Disk quota exceeded: '/home/vlad.smirnov/uhead_ft/x_llama31_8b'
14
+ [W1026 12:16:25.424616836 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
15
+ Traceback (most recent call last):
16
+ File "<frozen runpy>", line 198, in _run_module_as_main
17
+ File "<frozen runpy>", line 88, in _run_code
18
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 561, in <module>
19
+ main()
20
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 557, in main
21
+ (cmd_dump if a.cmd == "dump" else cmd_train)(a)
22
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 156, in cmd_dump
23
+ out = Path(a.out_dir); out.mkdir(parents=True, exist_ok=True)
24
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
25
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1116, in mkdir
26
+ os.mkdir(self, mode)
27
+ OSError: [Errno 122] Disk quota exceeded: '/home/vlad.smirnov/uhead_ft/x_llama31_8b'
28
+ [W1027 13:23:01.024341809 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
29
+ OSError: [Errno 122] Disk quota exceeded
30
+
31
+ During handling of the above exception, another exception occurred:
32
+
33
+ Traceback (most recent call last):
34
+ File "<frozen runpy>", line 198, in _run_module_as_main
35
+ File "<frozen runpy>", line 88, in _run_code
36
+ File "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py", line 561, in <module>
37
+ main()
38
+ File "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py", line 557, in main
39
+ (cmd_dump if a.cmd == "dump" else cmd_train)(a)
40
+ File "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py", line 172, in cmd_dump
41
+ provenance.write(out / "provenance.json", prov, status="running", resumed_chains=len(resumed[0]) if resumed else 0)
42
+ File "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/provenance.py", line 169, in write
43
+ Path(path).write_text(json.dumps(finish(prov, **extra), indent=1, default=str))
44
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1078, in write_text
45
+ with self.open(mode='w', encoding=encoding, errors=errors, newline=newline) as f:
46
+ OSError: [Errno 122] Disk quota exceeded
47
+ [W1030 12:14:53.010595942 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
48
+ `torch_dtype` is deprecated! Use `dtype` instead!
49
+
50
+ Loading uncertainty head from /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/uhead_claim_hs_Qwen3-8B
51
+ 2026-10-30 15:04:24,341 INFO luh.feature_extractors.combined: Loading feature extractor: luh.feature_extractors.basic_hidden_states
52
+ 2026-10-30 15:04:24,341 INFO root: Feature size: 4096
53
+ 2026-10-30 15:04:24,370 INFO root: Total number of parameters 8140801
54
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
55
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
56
+ 2026-10-30 15:04:25,894 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing hidden states generator for Qwen/Qwen3-8B
57
+ 2026-10-30 15:04:25,894 INFO speculators.data_generation.vllm_hidden_states_generator: Tensor parallel size: 1
58
+ 2026-10-30 15:04:26,109 INFO speculators.data_generation.vllm_hidden_states_generator: Model has 36 layers
59
+ 2026-10-30 15:04:26,109 INFO speculators.data_generation.vllm_hidden_states_generator: Using specified layers: [35]
60
+ The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
61
+ INFO 10-30 15:04:26 [model.py:541] Resolved architecture: Qwen3ForCausalLM
62
+ INFO 10-30 15:04:26 [model.py:1561] Using max model len 8192
63
+ INFO 10-30 15:04:26 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
64
+ INFO 10-30 15:04:26 [vllm.py:624] Asynchronous scheduling is enabled.
65
+ WARNING 10-30 15:04:26 [vllm.py:662] Enforce eager set, overriding optimization level to -O0
66
+ INFO 10-30 15:04:26 [vllm.py:762] Cudagraph is disabled under eager mode
67
+ 2026-10-30 15:04:26,115 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing executor...
68
+ WARNING 10-30 15:04:26 [multiproc_executor.py:910] Reducing Torch parallelism from 32 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
69
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
70
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
71
+ INFO 10-30 15:04:30 [worker_base.py:283] Injected <class 'speculators.data_generation.custom_worker.HiddenStatesWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_get_captured_states', '_reset_capture', '_set_request_metadata', '_setup_hidden_states_capture', '_store_captured_states']
72
+ INFO 10-30 15:04:30 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:33239 backend=nccl
73
+ INFO 10-30 15:04:30 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
74
+ (Worker pid=2380835) INFO 10-30 15:04:30 [gpu_model_runner.py:4033] Starting to load model Qwen/Qwen3-8B...
75
+ (Worker pid=2380835) INFO 10-30 15:04:31 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
76
+ (Worker pid=2380835)
77
+ (Worker pid=2380835)
78
+ (Worker pid=2380835)
79
+ (Worker pid=2380835)
80
+ (Worker pid=2380835)
81
+ (Worker pid=2380835)
82
+ (Worker pid=2380835)
83
+ (Worker pid=2380835)
84
+ (Worker pid=2380835) INFO 10-30 15:04:34 [default_loader.py:291] Loading weights took 3.08 seconds
85
+ (Worker pid=2380835) INFO 10-30 15:04:34 [gpu_model_runner.py:4130] Model loading took 15.27 GiB memory and 3.421976 seconds
86
+ 2026-10-30 15:04:34,890 INFO speculators.data_generation.vllm_hidden_states_generator: Setting up hidden states capture...
87
+ (Worker pid=2380835) [2026-10-30 15:04:34] INFO custom_worker.py:136: Hidden states capture setup complete for layers [35]
88
+ 2026-10-30 15:04:34,891 INFO speculators.data_generation.vllm_hidden_states_generator: Creating scheduler...
89
+ 2026-10-30 15:04:35,130 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing KV cache on all workers...
90
+ 2026-10-30 15:04:35,340 INFO rl_controller.feature_provider: VLLMFeatureProvider initialized: model=Qwen/Qwen3-8B, uhead_layers=[35], gen_layers=[35], head_dim=512
91
+ [dump] 8/10888 chains, 0 failed, 2787 rows, 0.0 min
92
+ (Worker pid=2380835) INFO 10-30 15:05:59 [multiproc_executor.py:730] Parent process exited, terminating worker
93
+ (Worker pid=2380835) INFO 10-30 15:05:59 [multiproc_executor.py:774] WorkerProc shutting down.
94
+ [W1030 15:06:00.665067346 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
95
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked shared_memory objects to clean up at shutdown
96
+ warnings.warn('resource_tracker: There appear to be %d '
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/base/uhead8b_selector.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/gpools/qwen3_1.7b/uhead8b_dump.log ADDED
@@ -0,0 +1,216 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Traceback (most recent call last):
2
+ File "<frozen runpy>", line 198, in _run_module_as_main
3
+ File "<frozen runpy>", line 88, in _run_code
4
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 532, in <module>
5
+ main()
6
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 528, in main
7
+ (cmd_dump if a.cmd == "dump" else cmd_train)(a)
8
+ File "/mnt/data/users/vlad.smirnov/tmp/wt_62aeb39d/rl_controller/scripts/uhead_finetune.py", line 137, in cmd_dump
9
+ out = Path(a.out_dir); out.mkdir(parents=True, exist_ok=True)
10
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
11
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1116, in mkdir
12
+ os.mkdir(self, mode)
13
+ OSError: [Errno 122] Disk quota exceeded: '/home/vlad.smirnov/uhead_ft/x_qwen3_1.7b'
14
+ [W1026 12:16:10.338675615 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
15
+ Traceback (most recent call last):
16
+ File "<frozen runpy>", line 198, in _run_module_as_main
17
+ File "<frozen runpy>", line 88, in _run_code
18
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 561, in <module>
19
+ main()
20
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 557, in main
21
+ (cmd_dump if a.cmd == "dump" else cmd_train)(a)
22
+ File "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/uhead_finetune.py", line 156, in cmd_dump
23
+ out = Path(a.out_dir); out.mkdir(parents=True, exist_ok=True)
24
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
25
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1116, in mkdir
26
+ os.mkdir(self, mode)
27
+ OSError: [Errno 122] Disk quota exceeded: '/home/vlad.smirnov/uhead_ft/x_qwen3_1.7b'
28
+ [W1027 13:22:58.226696067 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
29
+ `torch_dtype` is deprecated! Use `dtype` instead!
30
+
31
+ 2026-10-30 12:13:29,541 INFO matplotlib.font_manager: Failed to extract font properties from /usr/share/fonts/truetype/noto/NotoColorEmoji.ttf: Can not load face (unknown file format; error code 0x2)
32
+ 2026-10-30 12:13:29,794 INFO matplotlib.font_manager: generated new fontManager
33
+ Loading uncertainty head from /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/uhead_claim_hs_Qwen3-8B
34
+ 2026-10-30 12:13:30,200 INFO luh.feature_extractors.combined: Loading feature extractor: luh.feature_extractors.basic_hidden_states
35
+ 2026-10-30 12:13:30,200 INFO root: Feature size: 4096
36
+ 2026-10-30 12:13:30,264 INFO root: Total number of parameters 8140801
37
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
38
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
39
+ 2026-10-30 12:13:34,911 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing hidden states generator for Qwen/Qwen3-8B
40
+ 2026-10-30 12:13:34,911 INFO speculators.data_generation.vllm_hidden_states_generator: Tensor parallel size: 1
41
+ 2026-10-30 12:13:35,102 INFO speculators.data_generation.vllm_hidden_states_generator: Model has 36 layers
42
+ 2026-10-30 12:13:35,102 INFO speculators.data_generation.vllm_hidden_states_generator: Using specified layers: [35]
43
+ The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
44
+ INFO 10-30 12:13:35 [model.py:541] Resolved architecture: Qwen3ForCausalLM
45
+ INFO 10-30 12:13:35 [model.py:1561] Using max model len 8192
46
+ INFO 10-30 12:13:35 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
47
+ INFO 10-30 12:13:35 [vllm.py:624] Asynchronous scheduling is enabled.
48
+ WARNING 10-30 12:13:35 [vllm.py:662] Enforce eager set, overriding optimization level to -O0
49
+ INFO 10-30 12:13:35 [vllm.py:762] Cudagraph is disabled under eager mode
50
+ 2026-10-30 12:13:35,121 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing executor...
51
+ WARNING 10-30 12:13:35 [multiproc_executor.py:910] Reducing Torch parallelism from 32 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
52
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
53
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
54
+ INFO 10-30 12:13:39 [worker_base.py:283] Injected <class 'speculators.data_generation.custom_worker.HiddenStatesWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_get_captured_states', '_reset_capture', '_set_request_metadata', '_setup_hidden_states_capture', '_store_captured_states']
55
+ INFO 10-30 12:13:39 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:45877 backend=nccl
56
+ INFO 10-30 12:13:39 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
57
+ (Worker pid=2299516) INFO 10-30 12:13:40 [gpu_model_runner.py:4033] Starting to load model Qwen/Qwen3-8B...
58
+ (Worker pid=2299516) INFO 10-30 12:13:42 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
59
+ (Worker pid=2299516)
60
+ (Worker pid=2299516)
61
+ (Worker pid=2299516)
62
+ (Worker pid=2299516)
63
+ (Worker pid=2299516)
64
+ (Worker pid=2299516)
65
+ (Worker pid=2299516)
66
+ (Worker pid=2299516)
67
+ (Worker pid=2299516) INFO 10-30 12:13:45 [default_loader.py:291] Loading weights took 3.07 seconds
68
+ (Worker pid=2299516) INFO 10-30 12:13:46 [gpu_model_runner.py:4130] Model loading took 15.27 GiB memory and 4.398684 seconds
69
+ 2026-10-30 12:13:46,013 INFO speculators.data_generation.vllm_hidden_states_generator: Setting up hidden states capture...
70
+ (Worker pid=2299516) [2026-10-30 12:13:46] INFO custom_worker.py:136: Hidden states capture setup complete for layers [35]
71
+ 2026-10-30 12:13:46,014 INFO speculators.data_generation.vllm_hidden_states_generator: Creating scheduler...
72
+ 2026-10-30 12:13:46,261 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing KV cache on all workers...
73
+ 2026-10-30 12:13:46,595 INFO rl_controller.feature_provider: VLLMFeatureProvider initialized: model=Qwen/Qwen3-8B, uhead_layers=[35], gen_layers=[35], head_dim=512
74
+ [dump] 8/10888 chains, 0 failed, 3662 rows, 0.0 min
75
+ Traceback (most recent call last):
76
+ File "<frozen runpy>", line 198, in _run_module_as_main
77
+ File "<f(Worker pid=2299516) INFO 10-30 12:14:47 [multiproc_executor.py:730] Parent process exited, terminating worker
78
+ (Worker pid=2299516) INFO 10-30 12:14:47 [multiproc_executor.py:774] WorkerProc shutting down.
79
+ [W1030 12:14:48.040547559 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
80
+ [W1030 12:14:50.888895881 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
81
+ [dump] resuming: 8 chains kept, 10880 to go
82
+ `torch_dtype` is deprecated! Use `dtype` instead!
83
+
84
+ Loading uncertainty head from /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/uhead_claim_hs_Qwen3-8B
85
+ 2026-10-30 12:57:04,076 INFO luh.feature_extractors.combined: Loading feature extractor: luh.feature_extractors.basic_hidden_states
86
+ 2026-10-30 12:57:04,076 INFO root: Feature size: 4096
87
+ 2026-10-30 12:57:04,105 INFO root: Total number of parameters 8140801
88
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
89
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
90
+ 2026-10-30 12:57:08,190 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing hidden states generator for Qwen/Qwen3-8B
91
+ 2026-10-30 12:57:08,190 INFO speculators.data_generation.vllm_hidden_states_generator: Tensor parallel size: 1
92
+ 2026-10-30 12:57:08,383 INFO speculators.data_generation.vllm_hidden_states_generator: Model has 36 layers
93
+ 2026-10-30 12:57:08,383 INFO speculators.data_generation.vllm_hidden_states_generator: Using specified layers: [35]
94
+ The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
95
+ INFO 10-30 12:57:08 [model.py:541] Resolved architecture: Qwen3ForCausalLM
96
+ INFO 10-30 12:57:08 [model.py:1561] Using max model len 8192
97
+ INFO 10-30 12:57:08 [scheduler.py:226] Chunked prefill is enabled with max_num_batched_tokens=8192.
98
+ INFO 10-30 12:57:08 [vllm.py:624] Asynchronous scheduling is enabled.
99
+ WARNING 10-30 12:57:08 [vllm.py:662] Enforce eager set, overriding optimization level to -O0
100
+ INFO 10-30 12:57:08 [vllm.py:762] Cudagraph is disabled under eager mode
101
+ 2026-10-30 12:57:08,396 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing executor...
102
+ WARNING 10-30 12:57:08 [multiproc_executor.py:910] Reducing Torch parallelism from 32 threads to 1 to avoid unnecessary CPU contention. Set OMP_NUM_THREADS in the external environment to tune this value as needed.
103
+ /mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/site-packages/speculators/models/mlp.py:9: UserWarning: Field name "torch_dtype" in "MLPSpeculatorConfig" shadows an attribute in parent "SpeculatorModelConfig"
104
+ class MLPSpeculatorConfig(SpeculatorModelConfig):
105
+ INFO 10-30 12:57:12 [worker_base.py:283] Injected <class 'speculators.data_generation.custom_worker.HiddenStatesWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_get_captured_states', '_reset_capture', '_set_request_metadata', '_setup_hidden_states_capture', '_store_captured_states']
106
+ INFO 10-30 12:57:12 [parallel_state.py:1212] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:43415 backend=nccl
107
+ INFO 10-30 12:57:13 [parallel_state.py:1423] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A
108
+ (Worker pid=2342353) INFO 10-30 12:57:13 [gpu_model_runner.py:4033] Starting to load model Qwen/Qwen3-8B...
109
+ (Worker pid=2342353) INFO 10-30 12:57:14 [cuda.py:364] Using FLASH_ATTN attention backend out of potential backends: ('FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION')
110
+ (Worker pid=2342353)
111
+ (Worker pid=2342353)
112
+ (Worker pid=2342353)
113
+ (Worker pid=2342353)
114
+ (Worker pid=2342353)
115
+ (Worker pid=2342353)
116
+ (Worker pid=2342353)
117
+ (Worker pid=2342353)
118
+ (Worker pid=2342353) INFO 10-30 12:57:18 [default_loader.py:291] Loading weights took 3.08 seconds
119
+ (Worker pid=2342353) INFO 10-30 12:57:18 [gpu_model_runner.py:4130] Model loading took 15.27 GiB memory and 3.975952 seconds
120
+ 2026-10-30 12:57:18,555 INFO speculators.data_generation.vllm_hidden_states_generator: Setting up hidden states capture...
121
+ (Worker pid=2342353) [2026-10-30 12:57:18] INFO custom_worker.py:136: Hidden states capture setup complete for layers [35]
122
+ 2026-10-30 12:57:18,555 INFO speculators.data_generation.vllm_hidden_states_generator: Creating scheduler...
123
+ 2026-10-30 12:57:18,796 INFO speculators.data_generation.vllm_hidden_states_generator: Initializing KV cache on all workers...
124
+ 2026-10-30 12:57:19,016 INFO rl_controller.feature_provider: VLLMFeatureProvider initialized: model=Qwen/Qwen3-8B, uhead_layers=[35], gen_layers=[35], head_dim=512
125
+ [dump] 8/10880 chains, 0 failed, 8187 rows, 0.0 min
126
+ [dump] 208/10880 chains, 0 failed, 193870 rows, 1.3 min
127
+ 2026-10-30 12:59:52,956 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_44_7 not found in captured states. Available: ['req_44_0', 'req_44_1', 'req_44_2', 'req_44_3', 'req_44_4', 'req_44_5', 'req_44_6']); retrying one-at-a-time for 8 samples
128
+ 2026-10-30 13:00:47,616 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_58_7 not found in captured states. Available: ['req_58_0', 'req_58_1', 'req_58_2', 'req_58_3', 'req_58_4', 'req_58_5', 'req_58_6']); retrying one-at-a-time for 8 samples
129
+ [dump] 408/10880 chains, 0 failed, 458814 rows, 3.8 min
130
+ [dump] 608/10880 chains, 0 failed, 693363 rows, 5.8 min
131
+ 2026-10-30 13:03:12,459 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_93_7 not found in captured states. Available: ['req_93_0', 'req_93_1', 'req_93_2', 'req_93_3', 'req_93_4', 'req_93_5', 'req_93_6']); retrying one-at-a-time for 8 samples
132
+ 2026-10-30 13:05:20,285 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_123_6 not found in captured states. Available: ['req_123_0', 'req_123_1', 'req_123_2', 'req_123_3', 'req_123_4', 'req_123_5']); retrying one-at-a-time for 8 samples
133
+ [dump] 808/10880 chains, 0 failed, 952876 rows, 8.7 min
134
+ 2026-10-30 13:08:23,750 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_157_7 not found in captured states. Available: ['req_157_0', 'req_157_1', 'req_157_2', 'req_157_3', 'req_157_4', 'req_157_5', 'req_157_6']); retrying one-at-a-time for 8 samples
135
+ [dump] 1008/10880 chains, 0 failed, 1235013 rows, 11.5 min
136
+ 2026-10-30 13:08:50,769 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_166_7 not found in captured states. Available: ['req_166_0', 'req_166_1', 'req_166_2', 'req_166_3', 'req_166_4', 'req_166_5', 'req_166_6']); retrying one-at-a-time for 8 samples
137
+ 2026-10-30 13:09:42,531 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_179_7 not found in captured states. Available: ['req_179_0', 'req_179_1', 'req_179_2', 'req_179_3', 'req_179_4', 'req_179_5', 'req_179_6']); retrying one-at-a-time for 8 samples
138
+ [dump] 1208/10880 chains, 0 failed, 1480163 rows, 13.8 min
139
+ [dump] 1408/10880 chains, 0 failed, 1744563 rows, 16.2 min
140
+ 2026-10-30 13:14:47,255 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_241_7 not found in captured states. Available: ['req_241_0', 'req_241_1', 'req_241_2', 'req_241_3', 'req_241_4', 'req_241_5', 'req_241_6']); retrying one-at-a-time for 8 samples
141
+ 2026-10-30 13:15:20,690 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_253_7 not found in captured states. Available: ['req_253_0', 'req_253_1', 'req_253_2', 'req_253_3', 'req_253_4', 'req_253_5', 'req_253_6']); retrying one-at-a-time for 8 samples
142
+ [dump] 1608/10880 chains, 0 failed, 2032206 rows, 19.1 min
143
+ 2026-10-30 13:16:54,318 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_277_7 not found in captured states. Available: ['req_277_0', 'req_277_1', 'req_277_2', 'req_277_3', 'req_277_4', 'req_277_5', 'req_277_6']); retrying one-at-a-time for 8 samples
144
+ [dump] 1808/10880 chains, 0 failed, 2242764 rows, 20.9 min
145
+ 2026-10-30 13:20:19,971 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_329_7 not found in captured states. Available: ['req_329_0', 'req_329_1', 'req_329_2', 'req_329_3', 'req_329_4', 'req_329_5', 'req_329_6']); retrying one-at-a-time for 8 samples
146
+ [dump] 2008/10880 chains, 0 failed, 2493223 rows, 23.5 min
147
+ 2026-10-30 13:21:17,017 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_348_7 not found in captured states. Available: ['req_348_0', 'req_348_1', 'req_348_2', 'req_348_3', 'req_348_4', 'req_348_5', 'req_348_6']); retrying one-at-a-time for 8 samples
148
+ [dump] 2208/10880 chains, 0 failed, 2752804 rows, 26.3 min
149
+ [dump] 2408/10880 chains, 0 failed, 2963296 rows, 28.1 min
150
+ 2026-10-30 13:26:57,894 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_415_7 not found in captured states. Available: ['req_415_0', 'req_415_1', 'req_415_2', 'req_415_3', 'req_415_4', 'req_415_5', 'req_415_6']); retrying one-at-a-time for 8 samples
151
+ [dump] 2608/10880 chains, 0 failed, 3224989 rows, 30.6 min
152
+ 2026-10-30 13:29:41,600 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_452_7 not found in captured states. Available: ['req_452_0', 'req_452_1', 'req_452_2', 'req_452_3', 'req_452_4', 'req_452_5', 'req_452_6']); retrying one-at-a-time for 8 samples
153
+ [dump] 2808/10880 chains, 0 failed, 3464128 rows, 33.0 min
154
+ [dump] 3008/10880 chains, 0 failed, 3691801 rows, 34.8 min
155
+ [dump] 3208/10880 chains, 0 failed, 3959371 rows, 37.4 min
156
+ [dump] 3408/10880 chains, 0 failed, 4202228 rows, 39.7 min
157
+ [dump] 3608/10880 chains, 0 failed, 4427697 rows, 41.7 min
158
+ [dump] 3808/10880 chains, 0 failed, 4619548 rows, 43.2 min
159
+ 2026-10-30 13:41:52,316 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_607_7 not found in captured states. Available: ['req_607_0', 'req_607_1', 'req_607_2', 'req_607_3', 'req_607_4', 'req_607_5', 'req_607_6']); retrying one-at-a-time for 8 samples
160
+ [dump] 4008/10880 chains, 0 failed, 4859204 rows, 45.5 min
161
+ 2026-10-30 13:43:04,670 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_624_7 not found in captured states. Available: ['req_624_0', 'req_624_1', 'req_624_2', 'req_624_3', 'req_624_4', 'req_624_5', 'req_624_6']); retrying one-at-a-time for 8 samples
162
+ [dump] 4208/10880 chains, 0 failed, 5125426 rows, 48.1 min
163
+ [dump] 4408/10880 chains, 0 failed, 5366303 rows, 50.4 min
164
+ [dump] 4608/10880 chains, 0 failed, 5584567 rows, 52.5 min
165
+ [dump] 4808/10880 chains, 0 failed, 5868403 rows, 55.4 min
166
+ [dump] 5008/10880 chains, 0 failed, 6070209 rows, 57.2 min
167
+ 2026-10-30 13:56:02,306 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_769_6 not found in captured states. Available: ['req_769_0', 'req_769_1', 'req_769_2', 'req_769_3', 'req_769_4', 'req_769_5']); retrying one-at-a-time for 8 samples
168
+ [dump] 5208/10880 chains, 0 failed, 6334215 rows, 59.9 min
169
+ [dump] 5408/10880 chains, 0 failed, 6529935 rows, 61.6 min
170
+ [dump] 5608/10880 chains, 0 failed, 6759224 rows, 63.6 min
171
+ 2026-10-30 14:01:06,222 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_842_7 not found in captured states. Available: ['req_842_0', 'req_842_1', 'req_842_2', 'req_842_3', 'req_842_4', 'req_842_5', 'req_842_6']); retrying one-at-a-time for 8 samples
172
+ [dump] 5808/10880 chains, 0 failed, 6992870 rows, 65.9 min
173
+ [dump] 6008/10880 chains, 0 failed, 7168622 rows, 67.3 min
174
+ 2026-10-30 14:05:55,057 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_910_7 not found in captured states. Available: ['req_910_0', 'req_910_1', 'req_910_2', 'req_910_3', 'req_910_4', 'req_910_5', 'req_910_6']); retrying one-at-a-time for 8 samples
175
+ [dump] 6208/10880 chains, 0 failed, 7404667 rows, 69.6 min
176
+ [dump] 6408/10880 chains, 0 failed, 7598220 rows, 71.1 min
177
+ 2026-10-30 14:10:13,538 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_970_7 not found in captured states. Available: ['req_970_0', 'req_970_1', 'req_970_2', 'req_970_3', 'req_970_4', 'req_970_5', 'req_970_6']); retrying one-at-a-time for 8 samples
178
+ [dump] 6608/10880 chains, 0 failed, 7907318 rows, 74.3 min
179
+ 2026-10-30 14:12:42,487 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1001_7 not found in captured states. Available: ['req_1001_0', 'req_1001_1', 'req_1001_2', 'req_1001_3', 'req_1001_4', 'req_1001_5', 'req_1001_6']); retrying one-at-a-time for 8 samples
180
+ [dump] 6808/10880 chains, 0 failed, 8135088 rows, 76.5 min
181
+ [dump] 7008/10880 chains, 0 failed, 8409067 rows, 79.3 min
182
+ [dump] 7208/10880 chains, 0 failed, 8687924 rows, 82.0 min
183
+ [dump] 7408/10880 chains, 0 failed, 8913503 rows, 84.1 min
184
+ [dump] 7608/10880 chains, 0 failed, 9159310 rows, 86.4 min
185
+ 2026-10-30 14:26:16,064 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1144_7 not found in captured states. Available: ['req_1144_0', 'req_1144_1', 'req_1144_2', 'req_1144_3', 'req_1144_4', 'req_1144_5', 'req_1144_6']); retrying one-at-a-time for 8 samples
186
+ [dump] 7808/10880 chains, 0 failed, 9442378 rows, 89.4 min
187
+ 2026-10-30 14:28:15,430 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1169_7 not found in captured states. Available: ['req_1169_0', 'req_1169_1', 'req_1169_2', 'req_1169_3', 'req_1169_4', 'req_1169_5', 'req_1169_6']); retrying one-at-a-time for 8 samples
188
+ [dump] 8008/10880 chains, 0 failed, 9688180 rows, 92.0 min
189
+ 2026-10-30 14:29:56,565 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1192_7 not found in captured states. Available: ['req_1192_0', 'req_1192_1', 'req_1192_2', 'req_1192_3', 'req_1192_4', 'req_1192_5', 'req_1192_6']); retrying one-at-a-time for 8 samples
190
+ [dump] 8208/10880 chains, 0 failed, 9928469 rows, 94.4 min
191
+ [dump] 8408/10880 chains, 0 failed, 10178414 rows, 96.8 min
192
+ [dump] 8608/10880 chains, 0 failed, 10365901 rows, 98.2 min
193
+ [dump] 8808/10880 chains, 0 failed, 10598698 rows, 100.2 min
194
+ 2026-10-30 14:38:20,077 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1305_7 not found in captured states. Available: ['req_1305_0', 'req_1305_1', 'req_1305_2', 'req_1305_3', 'req_1305_4', 'req_1305_5', 'req_1305_6']); retrying one-at-a-time for 8 samples
195
+ 2026-10-30 14:39:03,111 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1318_7 not found in captured states. Available: ['req_1318_0', 'req_1318_1', 'req_1318_2', 'req_1318_3', 'req_1318_4', 'req_1318_5', 'req_1318_6']); retrying one-at-a-time for 8 samples
196
+ [dump] 9008/10880 chains, 0 failed, 10826053 rows, 102.6 min
197
+ 2026-10-30 14:41:09,629 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1348_7 not found in captured states. Available: ['req_1348_0', 'req_1348_1', 'req_1348_2', 'req_1348_3', 'req_1348_4', 'req_1348_5', 'req_1348_6']); retrying one-at-a-time for 8 samples
198
+ [dump] 9208/10880 chains, 0 failed, 11116138 rows, 105.6 min
199
+ 2026-10-30 14:44:30,448 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1392_7 not found in captured states. Available: ['req_1392_0', 'req_1392_1', 'req_1392_2', 'req_1392_3', 'req_1392_4', 'req_1392_5', 'req_1392_6']); retrying one-at-a-time for 8 samples
200
+ [dump] 9408/10880 chains, 0 failed, 11343970 rows, 107.7 min
201
+ [dump] 9608/10880 chains, 0 failed, 11556580 rows, 109.5 min
202
+ 2026-10-30 14:48:49,083 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1449_7 not found in captured states. Available: ['req_1449_0', 'req_1449_1', 'req_1449_2', 'req_1449_3', 'req_1449_4', 'req_1449_5', 'req_1449_6']); retrying one-at-a-time for 8 samples
203
+ [dump] 9808/10880 chains, 0 failed, 11823375 rows, 112.4 min
204
+ [dump] 10008/10880 chains, 0 failed, 12040105 rows, 114.5 min
205
+ [dump] 10208/10880 chains, 0 failed, 12303878 rows, 116.9 min
206
+ 2026-10-30 14:56:13,916 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1524_7 not found in captured states. Available: ['req_1524_0', 'req_1524_1', 'req_1524_2', 'req_1524_3', 'req_1524_4', 'req_1524_5', 'req_1524_6']); retrying one-at-a-time for 8 samples
207
+ [dump] 10408/10880 chains, 0 failed, 12615609 rows, 120.6 min
208
+ [dump] 10608/10880 chains, 0 failed, 12879674 rows, 123.3 min
209
+ [dump] 10808/10880 chains, 0 failed, 13113497 rows, 125.3 min
210
+ 2026-10-30 15:03:34,780 WARNING rl_controller.feature_provider: Batch hidden-state generation failed (RuntimeError: Request req_1600_7 not found in captured states. Available: ['req_1600_0', 'req_1600_1', 'req_1600_2', 'req_1600_3', 'req_1600_4', 'req_1600_5', 'req_1600_6']); retrying one-at-a-time for 8 samples
211
+ [dump] 10880/10880 chains, 0 failed, 13234818 rows, 126.8 min
212
+ [dump] done
213
+ (Worker pid=2342353) INFO 10-30 15:04:11 [multiproc_executor.py:730] Parent process exited, terminating worker
214
+ (Worker pid=2342353) INFO 10-30 15:04:11 [multiproc_executor.py:774] WorkerProc shutting down.
215
+ [W1030 15:04:12.922310429 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
216
+ [W1030 15:04:14.577706909 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/laneOWN_llama.log ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ 2026-10-31 14:18:27 lane OWN start a4c8898d Lane OWN: REPO= and SKIP_CTRL= to run the GPU part on another host bases: llama31_8b
2
+ 14:54:21 llama31_8b dump ok
3
+ 14:58:01 llama31_8b probes: own_L16 own_L24 own_L32
4
+ LANE_OWN_DONE
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/laneUX.log ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2026-10-30 12:12:32 lane UX start 28564408 Lanes UO/UX: DUMP_ROOT to place dumps outside a full home quota
2
+ qwen3_1.7b dump FAILED
3
+ File "<frozen runpy>", line 198, in _run_module_as_main
4
+ File "<f(Worker pid=2299516) INFO 10-30 12:14:47 [multiproc_executor.py:730] Parent process exited, terminating worker
5
+ (Worker pid=2299516) INFO 10-30 12:14:47 [multiproc_executor.py:774] WorkerProc shutting down.
6
+ [W1030 12:14:48.040547559 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
7
+ [W1030 12:14:50.888895881 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
8
+ llama31_8b dump FAILED
9
+ Path(path).write_text(json.dumps(finish(prov, **extra), indent=1, default=str))
10
+ File "/mnt/data/users/vlad.smirnov/conda_envs/rl-uhead/lib/python3.11/pathlib.py", line 1078, in write_text
11
+ with self.open(mode='w', encoding=encoding, errors=errors, newline=newline) as f:
12
+ OSError: [Errno 122] Disk quota exceeded
13
+ [W1030 12:14:53.010595942 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
14
+ LANE_UX_DONE
15
+ 2026-10-30 12:56:41 lane UX start 28564408 Lanes UO/UX: DUMP_ROOT to place dumps outside a full home quota
16
+ 15:04:14 qwen3_1.7b dump ok
17
+ [scores] 10888/10888 chains scored -> workdir/gpools/qwen3_1.7b/base/uhead8b_selector.json
18
+ [W1030 15:04:17.603936050 AllocatorConfig.cpp:28] Warning: PYTORCH_CUDA_ALLOC_CONF is deprecated, use PYTORCH_ALLOC_CONF instead (function operator())
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/dump.log ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ `torch_dtype` is deprecated! Use `dtype` instead!
2
+
3
+ [dump-own] 1/10888 chains, 0 failed, 0.0 min
4
+ [dump-own] 501/10888 chains, 0 failed, 1.6 min
5
+ [dump-own] 1001/10888 chains, 0 failed, 3.1 min
6
+ [dump-own] 1501/10888 chains, 0 failed, 4.8 min
7
+ [dump-own] 2001/10888 chains, 0 failed, 6.3 min
8
+ [dump-own] 2501/10888 chains, 0 failed, 8.0 min
9
+ [dump-own] 3001/10888 chains, 0 failed, 9.5 min
10
+ [dump-own] 3501/10888 chains, 0 failed, 11.1 min
11
+ [dump-own] 4001/10888 chains, 0 failed, 12.7 min
12
+ [dump-own] 4501/10888 chains, 0 failed, 14.3 min
13
+ [dump-own] 5001/10888 chains, 0 failed, 15.9 min
14
+ [dump-own] 5501/10888 chains, 0 failed, 17.4 min
15
+ [dump-own] 6001/10888 chains, 0 failed, 18.8 min
16
+ [dump-own] 6501/10888 chains, 0 failed, 20.4 min
17
+ [dump-own] 7001/10888 chains, 0 failed, 22.1 min
18
+ [dump-own] 7501/10888 chains, 0 failed, 23.6 min
19
+ [dump-own] 8001/10888 chains, 0 failed, 25.4 min
20
+ [dump-own] 8501/10888 chains, 0 failed, 27.0 min
21
+ [dump-own] 9001/10888 chains, 0 failed, 28.7 min
22
+ [dump-own] 9501/10888 chains, 0 failed, 30.3 min
23
+ [dump-own] 10001/10888 chains, 0 failed, 32.0 min
24
+ [dump-own] 10501/10888 chains, 0 failed, 33.8 min
25
+ [dump-own] 10888/10888 chains, 0 failed, 35.3 min
26
+ [dump-own] done
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16.log ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10868 with prm targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:470: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.5935 val bce 0.5571
6
+ [train] epoch 2: train bce 0.5543 val bce 0.5398
7
+ [train] epoch 3: train bce 0.5434 val bce 0.5375
8
+ [train] fold 0: {'fold': 0, 'train': 7823, 'val': 863, 'test': 2182, 'best_val_bce': 0.5375459421891574, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.5952 val bce 0.5560
10
+ [train] epoch 2: train bce 0.5554 val bce 0.5368
11
+ [train] epoch 3: train bce 0.5443 val bce 0.5382
12
+ [train] fold 1: {'fold': 1, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.5367719986439523, 'minutes': 0.23}
13
+ [train] epoch 1: train bce 0.5922 val bce 0.5695
14
+ [train] epoch 2: train bce 0.5514 val bce 0.5492
15
+ [train] epoch 3: train bce 0.5419 val bce 0.5390
16
+ [train] fold 2: {'fold': 2, 'train': 7832, 'val': 862, 'test': 2174, 'best_val_bce': 0.539007445048978, 'minutes': 0.23}
17
+ [train] epoch 1: train bce 0.5916 val bce 0.5571
18
+ [train] epoch 2: train bce 0.5496 val bce 0.5410
19
+ [train] epoch 3: train bce 0.5401 val bce 0.5337
20
+ [train] fold 3: {'fold': 3, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.5336860008293679, 'minutes': 0.22}
21
+ [train] epoch 1: train bce 0.5908 val bce 0.5588
22
+ [train] epoch 2: train bce 0.5532 val bce 0.5517
23
+ [train] epoch 3: train bce 0.5428 val bce 0.5410
24
+ [train] fold 4: {'fold': 4, 'train': 7836, 'val': 864, 'test': 2168, 'best_val_bce': 0.5409642726168475, 'minutes': 0.22}
25
+ [train] report: {
26
+ "target": "prm",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7823,
32
+ "val": 863,
33
+ "test": 2182,
34
+ "best_val_bce": 0.5375459421891574,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7835,
40
+ "val": 861,
41
+ "test": 2172,
42
+ "best_val_bce": 0.5367719986439523,
43
+ "minutes": 0.23
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7832,
48
+ "val": 862,
49
+ "test": 2174,
50
+ "best_val_bce": 0.539007445048978,
51
+ "minutes": 0.23
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7835,
56
+ "val": 861,
57
+ "test": 2172,
58
+ "best_val_bce": 0.5336860008293679,
59
+ "minutes": 0.22
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7836,
64
+ "val": 864,
65
+ "test": 2168,
66
+ "best_val_bce": 0.5409642726168475,
67
+ "minutes": 0.22
68
+ }
69
+ ],
70
+ "n_chains": 10868,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.681653316120757
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8289040506284979,
77
+ "chain_auroc_last": 0.7999046848218259,
78
+ "chain_auroc_mean": 0.8261376424628907,
79
+ "within_auc_min": 0.6519654377880184,
80
+ "n_mixed_problems": 775
81
+ },
82
+ "prm_teacher": {
83
+ "chain_auroc_min": 0.8953517750389156,
84
+ "chain_auroc_last": 0.8799120363529703,
85
+ "chain_auroc_mean": 0.884410459163078,
86
+ "within_auc_min": 0.8025986943164363,
87
+ "n_mixed_problems": 775
88
+ }
89
+ }
90
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:54:24+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "prm",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_L16"
23
+ ],
24
+ "git": {
25
+ "commit": "a4c8898d22a77eca179002ef8c56cdd240e89611",
26
+ "branch": "HEAD",
27
+ "subject": "Lane OWN: REPO= and SKIP_CTRL= to run the GPU part on another host",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "prm",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T14:55:35+04:00",
86
+ "wall_secs": 70.7,
87
+ "n_chains": 10868
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/report.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "prm",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7823,
8
+ "val": 863,
9
+ "test": 2182,
10
+ "best_val_bce": 0.5375459421891574,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7835,
16
+ "val": 861,
17
+ "test": 2172,
18
+ "best_val_bce": 0.5367719986439523,
19
+ "minutes": 0.23
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7832,
24
+ "val": 862,
25
+ "test": 2174,
26
+ "best_val_bce": 0.539007445048978,
27
+ "minutes": 0.23
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7835,
32
+ "val": 861,
33
+ "test": 2172,
34
+ "best_val_bce": 0.5336860008293679,
35
+ "minutes": 0.22
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7836,
40
+ "val": 864,
41
+ "test": 2168,
42
+ "best_val_bce": 0.5409642726168475,
43
+ "minutes": 0.22
44
+ }
45
+ ],
46
+ "n_chains": 10868,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.681653316120757
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8289040506284979,
53
+ "chain_auroc_last": 0.7999046848218259,
54
+ "chain_auroc_mean": 0.8261376424628907,
55
+ "within_auc_min": 0.6519654377880184,
56
+ "n_mixed_problems": 775
57
+ },
58
+ "prm_teacher": {
59
+ "chain_auroc_min": 0.8953517750389156,
60
+ "chain_auroc_last": 0.8799120363529703,
61
+ "chain_auroc_mean": 0.884410459163078,
62
+ "within_auc_min": 0.8025986943164363,
63
+ "n_mixed_problems": 775
64
+ }
65
+ }
66
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L16/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24.log ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10868 with prm targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:470: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.6011 val bce 0.5660
6
+ [train] epoch 2: train bce 0.5671 val bce 0.5520
7
+ [train] epoch 3: train bce 0.5536 val bce 0.5479
8
+ [train] fold 0: {'fold': 0, 'train': 7823, 'val': 863, 'test': 2182, 'best_val_bce': 0.5478663485753793, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.6015 val bce 0.5692
10
+ [train] epoch 2: train bce 0.5677 val bce 0.5513
11
+ [train] epoch 3: train bce 0.5543 val bce 0.5512
12
+ [train] fold 1: {'fold': 1, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.551238646181454, 'minutes': 0.23}
13
+ [train] epoch 1: train bce 0.5998 val bce 0.5821
14
+ [train] epoch 2: train bce 0.5641 val bce 0.5622
15
+ [train] epoch 3: train bce 0.5522 val bce 0.5472
16
+ [train] fold 2: {'fold': 2, 'train': 7832, 'val': 862, 'test': 2174, 'best_val_bce': 0.5471876986268768, 'minutes': 0.22}
17
+ [train] epoch 1: train bce 0.5999 val bce 0.5727
18
+ [train] epoch 2: train bce 0.5617 val bce 0.5526
19
+ [train] epoch 3: train bce 0.5490 val bce 0.5452
20
+ [train] fold 3: {'fold': 3, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.5452199948047884, 'minutes': 0.23}
21
+ [train] epoch 1: train bce 0.5985 val bce 0.5716
22
+ [train] epoch 2: train bce 0.5650 val bce 0.5632
23
+ [train] epoch 3: train bce 0.5517 val bce 0.5491
24
+ [train] fold 4: {'fold': 4, 'train': 7836, 'val': 864, 'test': 2168, 'best_val_bce': 0.5491441374108923, 'minutes': 0.23}
25
+ [train] report: {
26
+ "target": "prm",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7823,
32
+ "val": 863,
33
+ "test": 2182,
34
+ "best_val_bce": 0.5478663485753793,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7835,
40
+ "val": 861,
41
+ "test": 2172,
42
+ "best_val_bce": 0.551238646181454,
43
+ "minutes": 0.23
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7832,
48
+ "val": 862,
49
+ "test": 2174,
50
+ "best_val_bce": 0.5471876986268768,
51
+ "minutes": 0.22
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7835,
56
+ "val": 861,
57
+ "test": 2172,
58
+ "best_val_bce": 0.5452199948047884,
59
+ "minutes": 0.23
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7836,
64
+ "val": 864,
65
+ "test": 2168,
66
+ "best_val_bce": 0.5491441374108923,
67
+ "minutes": 0.23
68
+ }
69
+ ],
70
+ "n_chains": 10868,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.667128819644055
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8269698836872172,
77
+ "chain_auroc_last": 0.8023330511875426,
78
+ "chain_auroc_mean": 0.8237463637974624,
79
+ "within_auc_min": 0.6563563748079877,
80
+ "n_mixed_problems": 775
81
+ },
82
+ "prm_teacher": {
83
+ "chain_auroc_min": 0.8953517750389156,
84
+ "chain_auroc_last": 0.8799120363529703,
85
+ "chain_auroc_mean": 0.884410459163078,
86
+ "within_auc_min": 0.8025986943164363,
87
+ "n_mixed_problems": 775
88
+ }
89
+ }
90
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:55:37+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "prm",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_L24"
23
+ ],
24
+ "git": {
25
+ "commit": "a4c8898d22a77eca179002ef8c56cdd240e89611",
26
+ "branch": "HEAD",
27
+ "subject": "Lane OWN: REPO= and SKIP_CTRL= to run the GPU part on another host",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "prm",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T14:56:48+04:00",
86
+ "wall_secs": 71.2,
87
+ "n_chains": 10868
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/report.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "prm",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7823,
8
+ "val": 863,
9
+ "test": 2182,
10
+ "best_val_bce": 0.5478663485753793,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7835,
16
+ "val": 861,
17
+ "test": 2172,
18
+ "best_val_bce": 0.551238646181454,
19
+ "minutes": 0.23
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7832,
24
+ "val": 862,
25
+ "test": 2174,
26
+ "best_val_bce": 0.5471876986268768,
27
+ "minutes": 0.22
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7835,
32
+ "val": 861,
33
+ "test": 2172,
34
+ "best_val_bce": 0.5452199948047884,
35
+ "minutes": 0.23
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7836,
40
+ "val": 864,
41
+ "test": 2168,
42
+ "best_val_bce": 0.5491441374108923,
43
+ "minutes": 0.23
44
+ }
45
+ ],
46
+ "n_chains": 10868,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.667128819644055
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8269698836872172,
53
+ "chain_auroc_last": 0.8023330511875426,
54
+ "chain_auroc_mean": 0.8237463637974624,
55
+ "within_auc_min": 0.6563563748079877,
56
+ "n_mixed_problems": 775
57
+ },
58
+ "prm_teacher": {
59
+ "chain_auroc_min": 0.8953517750389156,
60
+ "chain_auroc_last": 0.8799120363529703,
61
+ "chain_auroc_mean": 0.884410459163078,
62
+ "within_auc_min": 0.8025986943164363,
63
+ "n_mixed_problems": 775
64
+ }
65
+ }
66
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L24/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32.log ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10868 with prm targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:470: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.6049 val bce 0.5684
6
+ [train] epoch 2: train bce 0.5745 val bce 0.5578
7
+ [train] epoch 3: train bce 0.5625 val bce 0.5490
8
+ [train] fold 0: {'fold': 0, 'train': 7823, 'val': 863, 'test': 2182, 'best_val_bce': 0.5490420723653779, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.6059 val bce 0.5752
10
+ [train] epoch 2: train bce 0.5756 val bce 0.5581
11
+ [train] epoch 3: train bce 0.5626 val bce 0.5538
12
+ [train] fold 1: {'fold': 1, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.553835202505352, 'minutes': 0.23}
13
+ [train] epoch 1: train bce 0.6026 val bce 0.5925
14
+ [train] epoch 2: train bce 0.5710 val bce 0.5739
15
+ [train] epoch 3: train bce 0.5603 val bce 0.5555
16
+ [train] fold 2: {'fold': 2, 'train': 7832, 'val': 862, 'test': 2174, 'best_val_bce': 0.5554942518190139, 'minutes': 0.23}
17
+ [train] epoch 1: train bce 0.6030 val bce 0.5780
18
+ [train] epoch 2: train bce 0.5691 val bce 0.5622
19
+ [train] epoch 3: train bce 0.5571 val bce 0.5581
20
+ [train] fold 3: {'fold': 3, 'train': 7835, 'val': 861, 'test': 2172, 'best_val_bce': 0.5580589989766874, 'minutes': 0.23}
21
+ [train] epoch 1: train bce 0.6013 val bce 0.5772
22
+ [train] epoch 2: train bce 0.5721 val bce 0.5779
23
+ [train] epoch 3: train bce 0.5587 val bce 0.5613
24
+ [train] fold 4: {'fold': 4, 'train': 7836, 'val': 864, 'test': 2168, 'best_val_bce': 0.5613067178004423, 'minutes': 0.23}
25
+ [train] report: {
26
+ "target": "prm",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7823,
32
+ "val": 863,
33
+ "test": 2182,
34
+ "best_val_bce": 0.5490420723653779,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7835,
40
+ "val": 861,
41
+ "test": 2172,
42
+ "best_val_bce": 0.553835202505352,
43
+ "minutes": 0.23
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7832,
48
+ "val": 862,
49
+ "test": 2174,
50
+ "best_val_bce": 0.5554942518190139,
51
+ "minutes": 0.23
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7835,
56
+ "val": 861,
57
+ "test": 2172,
58
+ "best_val_bce": 0.5580589989766874,
59
+ "minutes": 0.23
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7836,
64
+ "val": 864,
65
+ "test": 2168,
66
+ "best_val_bce": 0.5613067178004423,
67
+ "minutes": 0.23
68
+ }
69
+ ],
70
+ "n_chains": 10868,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.6508170989564003
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8226226889898656,
77
+ "chain_auroc_last": 0.8075478399598666,
78
+ "chain_auroc_mean": 0.8232405916565072,
79
+ "within_auc_min": 0.6605591397849462,
80
+ "n_mixed_problems": 775
81
+ },
82
+ "prm_teacher": {
83
+ "chain_auroc_min": 0.8953517750389156,
84
+ "chain_auroc_last": 0.8799120363529703,
85
+ "chain_auroc_mean": 0.884410459163078,
86
+ "within_auc_min": 0.8025986943164363,
87
+ "n_mixed_problems": 775
88
+ }
89
+ }
90
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:56:49+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "prm",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_L32"
23
+ ],
24
+ "git": {
25
+ "commit": "a4c8898d22a77eca179002ef8c56cdd240e89611",
26
+ "branch": "HEAD",
27
+ "subject": "Lane OWN: REPO= and SKIP_CTRL= to run the GPU part on another host",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "prm",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T14:58:01+04:00",
86
+ "wall_secs": 71.5,
87
+ "n_chains": 10868
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/report.json ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "prm",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7823,
8
+ "val": 863,
9
+ "test": 2182,
10
+ "best_val_bce": 0.5490420723653779,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7835,
16
+ "val": 861,
17
+ "test": 2172,
18
+ "best_val_bce": 0.553835202505352,
19
+ "minutes": 0.23
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7832,
24
+ "val": 862,
25
+ "test": 2174,
26
+ "best_val_bce": 0.5554942518190139,
27
+ "minutes": 0.23
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7835,
32
+ "val": 861,
33
+ "test": 2172,
34
+ "best_val_bce": 0.5580589989766874,
35
+ "minutes": 0.23
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7836,
40
+ "val": 864,
41
+ "test": 2168,
42
+ "best_val_bce": 0.5613067178004423,
43
+ "minutes": 0.23
44
+ }
45
+ ],
46
+ "n_chains": 10868,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.6508170989564003
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8226226889898656,
53
+ "chain_auroc_last": 0.8075478399598666,
54
+ "chain_auroc_mean": 0.8232405916565072,
55
+ "within_auc_min": 0.6605591397849462,
56
+ "n_mixed_problems": 775
57
+ },
58
+ "prm_teacher": {
59
+ "chain_auroc_min": 0.8953517750389156,
60
+ "chain_auroc_last": 0.8799120363529703,
61
+ "chain_auroc_mean": 0.884410459163078,
62
+ "within_auc_min": 0.8025986943164363,
63
+ "n_mixed_problems": 775
64
+ }
65
+ }
66
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_L32/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16.log ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10888 with outcome targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:476: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.4361 val bce 0.3911
6
+ [train] epoch 2: train bce 0.3986 val bce 0.3828
7
+ [train] epoch 3: train bce 0.3898 val bce 0.3767
8
+ [train] fold 0: {'fold': 0, 'train': 7840, 'val': 864, 'test': 2184, 'best_val_bce': 0.376730606137432, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.4326 val bce 0.3911
10
+ [train] epoch 2: train bce 0.3940 val bce 0.3890
11
+ [train] epoch 3: train bce 0.3863 val bce 0.3889
12
+ [train] fold 1: {'fold': 1, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.3889442651453201, 'minutes': 0.22}
13
+ [train] epoch 1: train bce 0.4141 val bce 0.4660
14
+ [train] epoch 2: train bce 0.3759 val bce 0.4570
15
+ [train] epoch 3: train bce 0.3678 val bce 0.4501
16
+ [train] fold 2: {'fold': 2, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.45011325512344613, 'minutes': 0.23}
17
+ [train] epoch 1: train bce 0.4243 val bce 0.4330
18
+ [train] epoch 2: train bce 0.3854 val bce 0.4287
19
+ [train] epoch 3: train bce 0.3774 val bce 0.4306
20
+ [train] fold 3: {'fold': 3, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.4286661611927542, 'minutes': 0.23}
21
+ [train] epoch 1: train bce 0.4218 val bce 0.4647
22
+ [train] epoch 2: train bce 0.3836 val bce 0.4646
23
+ [train] epoch 3: train bce 0.3742 val bce 0.4761
24
+ [train] fold 4: {'fold': 4, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.4646486737067859, 'minutes': 0.23}
25
+ [train] report: {
26
+ "target": "outcome",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7840,
32
+ "val": 864,
33
+ "test": 2184,
34
+ "best_val_bce": 0.376730606137432,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.3889442651453201,
43
+ "minutes": 0.22
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7848,
48
+ "val": 864,
49
+ "test": 2176,
50
+ "best_val_bce": 0.45011325512344613,
51
+ "minutes": 0.23
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7848,
56
+ "val": 864,
57
+ "test": 2176,
58
+ "best_val_bce": 0.4286661611927542,
59
+ "minutes": 0.23
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7848,
64
+ "val": 864,
65
+ "test": 2176,
66
+ "best_val_bce": 0.4646486737067859,
67
+ "minutes": 0.23
68
+ }
69
+ ],
70
+ "n_chains": 10888,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.5518319384372918
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8461816323371188,
77
+ "chain_auroc_last": 0.8464377807862242,
78
+ "chain_auroc_mean": 0.8489497364908177,
79
+ "within_auc_min": 0.6708057191948945,
80
+ "n_mixed_problems": 776
81
+ }
82
+ }
83
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:00:58+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "outcome",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_out_L16"
23
+ ],
24
+ "git": {
25
+ "commit": "121c39851c92762f0d09d8f7a56043aa697c4210",
26
+ "branch": "HEAD",
27
+ "subject": "uhead_finetune train --target outcome; lane OWN TARGET= switch",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "outcome",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T15:02:09+04:00",
86
+ "wall_secs": 71.1,
87
+ "n_chains": 10888
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/report.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "outcome",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7840,
8
+ "val": 864,
9
+ "test": 2184,
10
+ "best_val_bce": 0.376730606137432,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7848,
16
+ "val": 864,
17
+ "test": 2176,
18
+ "best_val_bce": 0.3889442651453201,
19
+ "minutes": 0.22
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7848,
24
+ "val": 864,
25
+ "test": 2176,
26
+ "best_val_bce": 0.45011325512344613,
27
+ "minutes": 0.23
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7848,
32
+ "val": 864,
33
+ "test": 2176,
34
+ "best_val_bce": 0.4286661611927542,
35
+ "minutes": 0.23
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.4646486737067859,
43
+ "minutes": 0.23
44
+ }
45
+ ],
46
+ "n_chains": 10888,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.5518319384372918
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8461816323371188,
53
+ "chain_auroc_last": 0.8464377807862242,
54
+ "chain_auroc_mean": 0.8489497364908177,
55
+ "within_auc_min": 0.6708057191948945,
56
+ "n_mixed_problems": 776
57
+ }
58
+ }
59
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L16/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24.log ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10888 with outcome targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:476: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.4392 val bce 0.3908
6
+ [train] epoch 2: train bce 0.4031 val bce 0.3856
7
+ [train] epoch 3: train bce 0.3947 val bce 0.3809
8
+ [train] fold 0: {'fold': 0, 'train': 7840, 'val': 864, 'test': 2184, 'best_val_bce': 0.38094971823113877, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.4360 val bce 0.3929
10
+ [train] epoch 2: train bce 0.3994 val bce 0.3898
11
+ [train] epoch 3: train bce 0.3924 val bce 0.3888
12
+ [train] fold 1: {'fold': 1, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.38876801869817845, 'minutes': 0.23}
13
+ [train] epoch 1: train bce 0.4179 val bce 0.4651
14
+ [train] epoch 2: train bce 0.3828 val bce 0.4572
15
+ [train] epoch 3: train bce 0.3744 val bce 0.4538
16
+ [train] fold 2: {'fold': 2, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.453779890354496, 'minutes': 0.23}
17
+ [train] epoch 1: train bce 0.4288 val bce 0.4370
18
+ [train] epoch 2: train bce 0.3918 val bce 0.4316
19
+ [train] epoch 3: train bce 0.3839 val bce 0.4320
20
+ [train] fold 3: {'fold': 3, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.4316302136311647, 'minutes': 0.23}
21
+ [train] epoch 1: train bce 0.4259 val bce 0.4627
22
+ [train] epoch 2: train bce 0.3890 val bce 0.4672
23
+ [train] epoch 3: train bce 0.3801 val bce 0.4812
24
+ [train] fold 4: {'fold': 4, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.46265577080209724, 'minutes': 0.23}
25
+ [train] report: {
26
+ "target": "outcome",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7840,
32
+ "val": 864,
33
+ "test": 2184,
34
+ "best_val_bce": 0.38094971823113877,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.38876801869817845,
43
+ "minutes": 0.23
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7848,
48
+ "val": 864,
49
+ "test": 2176,
50
+ "best_val_bce": 0.453779890354496,
51
+ "minutes": 0.23
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7848,
56
+ "val": 864,
57
+ "test": 2176,
58
+ "best_val_bce": 0.4316302136311647,
59
+ "minutes": 0.23
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7848,
64
+ "val": 864,
65
+ "test": 2176,
66
+ "best_val_bce": 0.46265577080209724,
67
+ "minutes": 0.23
68
+ }
69
+ ],
70
+ "n_chains": 10888,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.5391224792794732
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8379986622373962,
77
+ "chain_auroc_last": 0.8396319937305462,
78
+ "chain_auroc_mean": 0.8406535486921783,
79
+ "within_auc_min": 0.6629433603338242,
80
+ "n_mixed_problems": 776
81
+ }
82
+ }
83
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:02:10+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "outcome",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_out_L24"
23
+ ],
24
+ "git": {
25
+ "commit": "121c39851c92762f0d09d8f7a56043aa697c4210",
26
+ "branch": "HEAD",
27
+ "subject": "uhead_finetune train --target outcome; lane OWN TARGET= switch",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "outcome",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T15:03:22+04:00",
86
+ "wall_secs": 71.1,
87
+ "n_chains": 10888
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/report.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "outcome",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7840,
8
+ "val": 864,
9
+ "test": 2184,
10
+ "best_val_bce": 0.38094971823113877,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7848,
16
+ "val": 864,
17
+ "test": 2176,
18
+ "best_val_bce": 0.38876801869817845,
19
+ "minutes": 0.23
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7848,
24
+ "val": 864,
25
+ "test": 2176,
26
+ "best_val_bce": 0.453779890354496,
27
+ "minutes": 0.23
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7848,
32
+ "val": 864,
33
+ "test": 2176,
34
+ "best_val_bce": 0.4316302136311647,
35
+ "minutes": 0.23
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.46265577080209724,
43
+ "minutes": 0.23
44
+ }
45
+ ],
46
+ "n_chains": 10888,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.5391224792794732
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8379986622373962,
53
+ "chain_auroc_last": 0.8396319937305462,
54
+ "chain_auroc_mean": 0.8406535486921783,
55
+ "within_auc_min": 0.6629433603338242,
56
+ "n_mixed_problems": 776
57
+ }
58
+ }
59
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L24/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32.log ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [train] 10888 dumped chains, 10888 with outcome targets, 1361 problems, head raw_probe, device cuda:0
2
+ /mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py:476: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
3
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
4
+ acc_l += float(o[0]); acc_n += o[1]
5
+ [train] epoch 1: train bce 0.4422 val bce 0.3969
6
+ [train] epoch 2: train bce 0.4061 val bce 0.3895
7
+ [train] epoch 3: train bce 0.3966 val bce 0.3835
8
+ [train] fold 0: {'fold': 0, 'train': 7840, 'val': 864, 'test': 2184, 'best_val_bce': 0.38353117774876916, 'minutes': 0.24}
9
+ [train] epoch 1: train bce 0.4381 val bce 0.3981
10
+ [train] epoch 2: train bce 0.4033 val bce 0.3959
11
+ [train] epoch 3: train bce 0.3943 val bce 0.3974
12
+ [train] fold 1: {'fold': 1, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.39589007674290294, 'minutes': 0.23}
13
+ [train] epoch 1: train bce 0.4212 val bce 0.4734
14
+ [train] epoch 2: train bce 0.3868 val bce 0.4542
15
+ [train] epoch 3: train bce 0.3775 val bce 0.4558
16
+ [train] fold 2: {'fold': 2, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.4541876046107073, 'minutes': 0.22}
17
+ [train] epoch 1: train bce 0.4297 val bce 0.4445
18
+ [train] epoch 2: train bce 0.3952 val bce 0.4363
19
+ [train] epoch 3: train bce 0.3866 val bce 0.4316
20
+ [train] fold 3: {'fold': 3, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.4316076996006977, 'minutes': 0.22}
21
+ [train] epoch 1: train bce 0.4291 val bce 0.4684
22
+ [train] epoch 2: train bce 0.3922 val bce 0.4727
23
+ [train] epoch 3: train bce 0.3815 val bce 0.4784
24
+ [train] fold 4: {'fold': 4, 'train': 7848, 'val': 864, 'test': 2176, 'best_val_bce': 0.46843065794036687, 'minutes': 0.22}
25
+ [train] report: {
26
+ "target": "outcome",
27
+ "head": "raw_probe",
28
+ "folds": [
29
+ {
30
+ "fold": 0,
31
+ "train": 7840,
32
+ "val": 864,
33
+ "test": 2184,
34
+ "best_val_bce": 0.38353117774876916,
35
+ "minutes": 0.24
36
+ },
37
+ {
38
+ "fold": 1,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.39589007674290294,
43
+ "minutes": 0.23
44
+ },
45
+ {
46
+ "fold": 2,
47
+ "train": 7848,
48
+ "val": 864,
49
+ "test": 2176,
50
+ "best_val_bce": 0.4541876046107073,
51
+ "minutes": 0.22
52
+ },
53
+ {
54
+ "fold": 3,
55
+ "train": 7848,
56
+ "val": 864,
57
+ "test": 2176,
58
+ "best_val_bce": 0.4316076996006977,
59
+ "minutes": 0.22
60
+ },
61
+ {
62
+ "fold": 4,
63
+ "train": 7848,
64
+ "val": 864,
65
+ "test": 2176,
66
+ "best_val_bce": 0.46843065794036687,
67
+ "minutes": 0.22
68
+ }
69
+ ],
70
+ "n_chains": 10888,
71
+ "step_corr_with_target": {
72
+ "finetuned": 0.5279939453941649
73
+ },
74
+ "chain_metrics": {
75
+ "finetuned": {
76
+ "chain_auroc_min": 0.8324295121167629,
77
+ "chain_auroc_last": 0.8345028837698029,
78
+ "chain_auroc_mean": 0.8344212150848435,
79
+ "within_auc_min": 0.6670939187530682,
80
+ "n_mixed_problems": 776
81
+ }
82
+ }
83
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/provenance.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T15:03:23+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "train",
9
+ "--features-dir",
10
+ "workdir/own_ft/llama31_8b",
11
+ "--target",
12
+ "outcome",
13
+ "--prm-selector",
14
+ "workdir/gpools/llama31_8b/base/prm7b_selector.json",
15
+ "--pool-dir",
16
+ "workdir/gpools/llama31_8b/chatpool",
17
+ "--head",
18
+ "raw_probe",
19
+ "--raw-key",
20
+ "<redacted>",
21
+ "--out-dir",
22
+ "workdir/own_ft/llama31_8b/probe_out_L32"
23
+ ],
24
+ "git": {
25
+ "commit": "121c39851c92762f0d09d8f7a56043aa697c4210",
26
+ "branch": "HEAD",
27
+ "subject": "uhead_finetune train --target outcome; lane OWN TARGET= switch",
28
+ "dirty": false,
29
+ "dirty_files": []
30
+ },
31
+ "gpus": {
32
+ "visible": "1",
33
+ "devices": [
34
+ {
35
+ "index": "0",
36
+ "name": "NVIDIA RTX 5880 Ada Generation",
37
+ "memory": "49140 MiB",
38
+ "driver": "580.178.04",
39
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
40
+ },
41
+ {
42
+ "index": "1",
43
+ "name": "NVIDIA RTX 5880 Ada Generation",
44
+ "memory": "49140 MiB",
45
+ "driver": "580.178.04",
46
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
47
+ }
48
+ ]
49
+ },
50
+ "packages": {
51
+ "python": "3.11.15",
52
+ "torch": "2.9.1",
53
+ "vllm": "0.15.1",
54
+ "transformers": "4.57.6",
55
+ "speculators": "0.4.0.dev49",
56
+ "peft": null,
57
+ "accelerate": "1.13.0",
58
+ "numpy": "2.2.6",
59
+ "huggingface_hub": "0.36.2",
60
+ "openai": "2.32.0",
61
+ "wandb": "0.25.1",
62
+ "torch_cuda": "12.8"
63
+ },
64
+ "env": {
65
+ "CUDA_VISIBLE_DEVICES": "1",
66
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
67
+ "HF_HUB_OFFLINE": "1",
68
+ "TRANSFORMERS_OFFLINE": "1"
69
+ },
70
+ "models": [],
71
+ "files": {
72
+ "features": {
73
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/own_ft/llama31_8b/provenance.json",
74
+ "bytes": 2375,
75
+ "sha256": "042d49f5d1fa1079738c3b5a1902cc953b4dbc172ff84192faa79933c37b19af"
76
+ },
77
+ "prm_selector": {
78
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/base/prm7b_selector.json",
79
+ "bytes": 8984153,
80
+ "sha256": "8f09bc1045ea0657ecebaae58bfbfa0d86a47236e99907aa3d99ce5d7f33bcb0"
81
+ }
82
+ },
83
+ "target": "outcome",
84
+ "head": "raw_probe",
85
+ "finished_at": "2026-10-31T15:04:33+04:00",
86
+ "wall_secs": 70.3,
87
+ "n_chains": 10888
88
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/report.json ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "target": "outcome",
3
+ "head": "raw_probe",
4
+ "folds": [
5
+ {
6
+ "fold": 0,
7
+ "train": 7840,
8
+ "val": 864,
9
+ "test": 2184,
10
+ "best_val_bce": 0.38353117774876916,
11
+ "minutes": 0.24
12
+ },
13
+ {
14
+ "fold": 1,
15
+ "train": 7848,
16
+ "val": 864,
17
+ "test": 2176,
18
+ "best_val_bce": 0.39589007674290294,
19
+ "minutes": 0.23
20
+ },
21
+ {
22
+ "fold": 2,
23
+ "train": 7848,
24
+ "val": 864,
25
+ "test": 2176,
26
+ "best_val_bce": 0.4541876046107073,
27
+ "minutes": 0.22
28
+ },
29
+ {
30
+ "fold": 3,
31
+ "train": 7848,
32
+ "val": 864,
33
+ "test": 2176,
34
+ "best_val_bce": 0.4316076996006977,
35
+ "minutes": 0.22
36
+ },
37
+ {
38
+ "fold": 4,
39
+ "train": 7848,
40
+ "val": 864,
41
+ "test": 2176,
42
+ "best_val_bce": 0.46843065794036687,
43
+ "minutes": 0.22
44
+ }
45
+ ],
46
+ "n_chains": 10888,
47
+ "step_corr_with_target": {
48
+ "finetuned": 0.5279939453941649
49
+ },
50
+ "chain_metrics": {
51
+ "finetuned": {
52
+ "chain_auroc_min": 0.8324295121167629,
53
+ "chain_auroc_last": 0.8345028837698029,
54
+ "chain_auroc_mean": 0.8344212150848435,
55
+ "within_auc_min": 0.6670939187530682,
56
+ "n_mixed_problems": 776
57
+ }
58
+ }
59
+ }
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/probe_out_L32/scores.json ADDED
The diff for this file is too large to render. See raw diff
 
judge-imitation/artem1/llm-reasoning-rl-with-uhead-workdir/own_ft/llama31_8b/provenance.json ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "started_at": "2026-10-31T14:19:00+04:00",
3
+ "host": "ws006598",
4
+ "user": "vlad.smirnov",
5
+ "cwd": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead",
6
+ "argv": [
7
+ "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/rl_controller/scripts/uhead_finetune.py",
8
+ "dump-own",
9
+ "--pool-dir",
10
+ "workdir/gpools/llama31_8b/chatpool",
11
+ "--base-model",
12
+ "meta-llama/Llama-3.1-8B-Instruct",
13
+ "--template",
14
+ "auto",
15
+ "--out-dir",
16
+ "workdir/own_ft/llama31_8b",
17
+ "--resume"
18
+ ],
19
+ "git": {
20
+ "commit": "a4c8898d22a77eca179002ef8c56cdd240e89611",
21
+ "branch": "HEAD",
22
+ "subject": "Lane OWN: REPO= and SKIP_CTRL= to run the GPU part on another host",
23
+ "dirty": false,
24
+ "dirty_files": []
25
+ },
26
+ "gpus": {
27
+ "visible": "1",
28
+ "devices": [
29
+ {
30
+ "index": "0",
31
+ "name": "NVIDIA RTX 5880 Ada Generation",
32
+ "memory": "49140 MiB",
33
+ "driver": "580.178.04",
34
+ "uuid": "GPU-eb08932b-fbcd-41ce-e1e2-6fd7b02de161"
35
+ },
36
+ {
37
+ "index": "1",
38
+ "name": "NVIDIA RTX 5880 Ada Generation",
39
+ "memory": "49140 MiB",
40
+ "driver": "580.178.04",
41
+ "uuid": "GPU-cc1e8db2-7843-a8e9-8191-088167fb3499"
42
+ }
43
+ ]
44
+ },
45
+ "packages": {
46
+ "python": "3.11.15",
47
+ "torch": "2.9.1",
48
+ "vllm": "0.15.1",
49
+ "transformers": "4.57.6",
50
+ "speculators": "0.4.0.dev49",
51
+ "peft": null,
52
+ "accelerate": "1.13.0",
53
+ "numpy": "2.2.6",
54
+ "huggingface_hub": "0.36.2",
55
+ "openai": "2.32.0",
56
+ "wandb": "0.25.1",
57
+ "torch_cuda": "12.8"
58
+ },
59
+ "env": {
60
+ "CUDA_VISIBLE_DEVICES": "1",
61
+ "HF_HOME": "/mnt/data/users/vlad.smirnov/hf_cache",
62
+ "HF_HUB_OFFLINE": "1",
63
+ "TRANSFORMERS_OFFLINE": "1",
64
+ "RLC_STEP_FORMAT": "chat"
65
+ },
66
+ "models": [
67
+ {
68
+ "repo": "meta-llama/Llama-3.1-8B-Instruct",
69
+ "revision": "0e9e39f249a16976918f6564b8830bc894c89659",
70
+ "path": "/mnt/data/users/vlad.smirnov/hf_cache/hub/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f249a16976918f6564b8830bc894c89659"
71
+ }
72
+ ],
73
+ "files": {
74
+ "pool_config": {
75
+ "path": "/mnt/data/users/vlad.smirnov/llm-reasoning-rl-with-uhead/workdir/gpools/llama31_8b/chatpool/pool_config.json",
76
+ "bytes": 213,
77
+ "sha256": "2442b361015d3bdac6fd3c2c74e57151637e209015d0058bd595d0b955fd8d9e"
78
+ }
79
+ },
80
+ "chains": [
81
+ 0,
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 4,
86
+ 5,
87
+ 6,
88
+ 7
89
+ ],
90
+ "layers": [
91
+ 16,
92
+ 24,
93
+ 32
94
+ ],
95
+ "template": "auto",
96
+ "finished_at": "2026-10-31T14:54:20+04:00",
97
+ "wall_secs": 2120.0,
98
+ "status": "done",
99
+ "n_chains": 10888,
100
+ "n_failed": 0
101
+ }