Hugging Face's logo

Datasets:
YFanwang
/
Backup

Dataset card Files
xet
 Community
YFanwang commited on
Commit
3207859
·
verified ·
1 Parent(s): ce5848d

Upload SAGE-GRPO outputs unifiedreward_think_pairwise

Browse files
Files changed (6) hide show
  1. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/logs/pairwise_eval.log +56 -0
  2. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/pairwise_results.jsonl +0 -0
  3. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/summary.json +31 -0
  4. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/logs/pairwise_eval.log +244 -0
  5. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/pairwise_results.jsonl +0 -0
  6. outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/summary.json +31 -0
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/logs/pairwise_eval.log ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-05-21 22:53:32] DIR_A: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100
2
+ [2026-05-21 22:53:32] DIR_B: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-1-sage-qwen35-9b-logprob/videogen_val/step_000100
3
+ [2026-05-21 22:53:32] MODEL_A_NAME: yw002-1-sage-baseline
4
+ [2026-05-21 22:53:32] MODEL_B_NAME: yw006-1-sage-qwen35-9b-logprob
5
+ [2026-05-21 22:53:32] OUTPUT_DIR: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100
6
+ [2026-05-21 22:53:32] REWARD_MODEL_PATH: /root/ckpts/UnifiedReward-Think-qwen35-9b
7
+ [2026-05-21 22:53:32] INSTALL_DEPS: true
8
+ [2026-05-21 22:53:32] INSTALL_TORCH: auto
9
+ [2026-05-21 22:53:32] DEVICE_MAP: auto
10
+ [2026-05-21 22:53:32] torch is already available.
11
+ [2026-05-21 22:53:32] Python client/model deps are already available.
12
+ dir_a videos: 300
13
+ dir_b videos: 300
14
+ paired videos selected for this shard: 300
15
+ already completed: 0
16
+ remaining pairs: 300
17
+ reward_model_path: /root/ckpts/UnifiedReward-Think-qwen35-9b
18
+ Detected reward checkpoint: model_type=qwen3_5, architectures=['Qwen3_5ForConditionalGeneration']
19
+ Candidate model classes: Qwen3_5ForConditionalGeneration, AutoModelForImageTextToText
20
+ [transformers] The fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and https://github.com/Dao-AILab/causal-conv1d
21
+
22
+
23
+ {
24
+ "total": 300,
25
+ "success": 300,
26
+ "counts": {
27
+ "yw002-1-sage-baseline": 122,
28
+ "yw006-1-sage-qwen35-9b-logprob": 178,
29
+ "tie": 0,
30
+ "unknown": 0,
31
+ "failed": 0
32
+ },
33
+ "win_rate_including_ties": {
34
+ "yw002-1-sage-baseline": 0.4066666666666667,
35
+ "yw006-1-sage-qwen35-9b-logprob": 0.5933333333333334
36
+ },
37
+ "win_rate_decisive_only": {
38
+ "yw002-1-sage-baseline": 0.4066666666666667,
39
+ "yw006-1-sage-qwen35-9b-logprob": 0.5933333333333334
40
+ },
41
+ "latency_sec": {
42
+ "mean": 37.73654333333333,
43
+ "median": 37.2285,
44
+ "min": 2.334,
45
+ "max": 95.853
46
+ },
47
+ "dir_a": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100",
48
+ "dir_b": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-1-sage-qwen35-9b-logprob/videogen_val/step_000100",
49
+ "reward_model_path": "/root/ckpts/UnifiedReward-Think-qwen35-9b",
50
+ "output_jsonl": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/pairwise_results.jsonl",
51
+ "shard_count": 1,
52
+ "shard_index": 0
53
+ }
54
+ [2026-05-22 02:02:42] Done.
55
+ [2026-05-22 02:02:42] Results: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/pairwise_results.jsonl
56
+ [2026-05-22 02:02:42] Summary: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/summary.json
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/pairwise_results.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/summary.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "total": 300,
3
+ "success": 300,
4
+ "counts": {
5
+ "yw002-1-sage-baseline": 122,
6
+ "yw006-1-sage-qwen35-9b-logprob": 178,
7
+ "tie": 0,
8
+ "unknown": 0,
9
+ "failed": 0
10
+ },
11
+ "win_rate_including_ties": {
12
+ "yw002-1-sage-baseline": 0.4066666666666667,
13
+ "yw006-1-sage-qwen35-9b-logprob": 0.5933333333333334
14
+ },
15
+ "win_rate_decisive_only": {
16
+ "yw002-1-sage-baseline": 0.4066666666666667,
17
+ "yw006-1-sage-qwen35-9b-logprob": 0.5933333333333334
18
+ },
19
+ "latency_sec": {
20
+ "mean": 37.73654333333333,
21
+ "median": 37.2285,
22
+ "min": 2.334,
23
+ "max": 95.853
24
+ },
25
+ "dir_a": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100",
26
+ "dir_b": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-1-sage-qwen35-9b-logprob/videogen_val/step_000100",
27
+ "reward_model_path": "/root/ckpts/UnifiedReward-Think-qwen35-9b",
28
+ "output_jsonl": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-1-sage-qwen35-9b-logprob/step_000100/pairwise_results.jsonl",
29
+ "shard_count": 1,
30
+ "shard_index": 0
31
+ }
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/logs/pairwise_eval.log ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [2026-05-21 19:03:01] DIR_A: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100
2
+ [2026-05-21 19:03:01] DIR_B: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100
3
+ [2026-05-21 19:03:01] MODEL_A_NAME: yw002-1-sage-baseline
4
+ [2026-05-21 19:03:01] MODEL_B_NAME: yw006-4-sage-qwen35-9b-token-credit-video-adv-2
5
+ [2026-05-21 19:03:01] OUTPUT_DIR: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100
6
+ [2026-05-21 19:03:01] API_URL: http://localhost:8080
7
+ [2026-05-21 19:03:01] MODEL_PATH: CodeGoat24/UnifiedReward-Think-qwen3vl-8b
8
+ [2026-05-21 19:03:01] INSTALL_DEPS: true
9
+ [2026-05-21 19:03:01] Python client deps are already available.
10
+ [ERROR] server did not become ready within 3s: http://localhost:8080
11
+ [ERROR] vllm command was not found. Install vLLM first, or rerun with INSTALL_VLLM=true if this server can build/install it.
12
+ [2026-05-21 19:10:01] DIR_A: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100
13
+ [2026-05-21 19:10:01] DIR_B: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100
14
+ [2026-05-21 19:10:01] MODEL_A_NAME: yw002-1-sage-baseline
15
+ [2026-05-21 19:10:01] MODEL_B_NAME: yw006-4-sage-qwen35-9b-token-credit-video-adv-2
16
+ [2026-05-21 19:10:01] OUTPUT_DIR: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100
17
+ [2026-05-21 19:10:01] REWARD_MODEL_PATH: CodeGoat24/UnifiedReward-Think-qwen3vl-8b
18
+ [2026-05-21 19:10:01] INSTALL_DEPS: true
19
+ [2026-05-21 19:10:01] INSTALL_TORCH: auto
20
+ [2026-05-21 19:10:01] DEVICE_MAP: auto
21
+ [2026-05-21 19:10:01] torch is already available.
22
+ [2026-05-21 19:10:01] Python client/model deps are already available.
23
+ dir_a videos: 300
24
+ dir_b videos: 300
25
+ paired videos selected for this shard: 300
26
+ already completed: 0
27
+ remaining pairs: 300
28
+ reward_model_path: CodeGoat24/UnifiedReward-Think-qwen3vl-8b
29
+ Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
30
+
31
+
32
+
33
+ [2026-05-21 19:12:05] DIR_B: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100
34
+ [2026-05-21 19:12:05] MODEL_A_NAME: yw002-1-sage-baseline
35
+ [2026-05-21 19:12:05] MODEL_B_NAME: yw006-4-sage-qwen35-9b-token-credit-video-adv-2
36
+ [2026-05-21 19:12:05] OUTPUT_DIR: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100
37
+ [2026-05-21 19:12:05] REWARD_MODEL_PATH: /root/ckpts/UnifiedReward-Think-qwen35-9b
38
+ [2026-05-21 19:12:05] INSTALL_DEPS: true
39
+ [2026-05-21 19:12:05] INSTALL_TORCH: auto
40
+ [2026-05-21 19:12:05] DEVICE_MAP: auto
41
+ [2026-05-21 19:12:05] torch is already available.
42
+ [2026-05-21 19:12:05] Python client/model deps are already available.
43
+ dir_a videos: 300
44
+ dir_b videos: 300
45
+ paired videos selected for this shard: 300
46
+ already completed: 0
47
+ remaining pairs: 300
48
+ reward_model_path: /root/ckpts/UnifiedReward-Think-qwen35-9b
49
+ [transformers] You are using a model of type `qwen3_5` to instantiate a model of type `qwen3_vl`. This may be expected if you are loading a checkpoint that shares a subset of the architecture (e.g., loading a `sam2_video` checkpoint into `Sam2Model`), but is otherwise not supported and can yield errors. Please verify that the checkpoint is compatible with the model you are instantiating.
50
+
51
+ [transformers] Qwen3VLForConditionalGeneration LOAD REPORT from: /root/ckpts/UnifiedReward-Think-qwen35-9b
52
+ Key | Status |
53
+ -----------------------------------------------------------------------------------+------------+-----------------------------------------------------------------------------------------------
54
+ model.language_model.layers.{0...30}.linear_attn.dt_bias | UNEXPECTED |
55
+ model.language_model.layers.{0...30}.linear_attn.in_proj_b.weight | UNEXPECTED |
56
+ model.language_model.layers.{0...30}.linear_attn.conv1d.weight | UNEXPECTED |
57
+ model.language_model.layers.{0...30}.linear_attn.in_proj_a.weight | UNEXPECTED |
58
+ model.language_model.layers.{0...30}.linear_attn.A_log | UNEXPECTED |
59
+ model.language_model.layers.{0...30}.linear_attn.in_proj_z.weight | UNEXPECTED |
60
+ model.language_model.layers.{0...30}.linear_attn.out_proj.weight | UNEXPECTED |
61
+ model.language_model.layers.{0...30}.linear_attn.norm.weight | UNEXPECTED |
62
+ model.language_model.layers.{0...30}.linear_attn.in_proj_qkv.weight | UNEXPECTED |
63
+ model.language_model.layers.{0...30}.self_attn.q_norm.weight | MISSING |
64
+ model.language_model.layers.{0...30}.self_attn.o_proj.weight | MISSING |
65
+ model.language_model.layers.{0...30}.self_attn.k_norm.weight | MISSING |
66
+ model.language_model.layers.{0...30}.self_attn.q_proj.weight | MISSING |
67
+ model.language_model.layers.{0...30}.self_attn.k_proj.weight | MISSING |
68
+ model.language_model.layers.{0...30}.self_attn.v_proj.weight | MISSING |
69
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.q_proj.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([8192, 4096]) vs model:torch.Size([4096, 4096])
70
+
71
+ Notes:
72
+ - UNEXPECTED: can be ignored when loading from different task/architecture; not ok if you expect identical arch.
73
+ - MISSING: those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
74
+ - MISMATCH: ckpt weights were loaded, but they did not match the original empty weight shapes.
75
+ [transformers] You are using a model of type `qwen3_5` to instantiate a model of type `qwen2_5_vl`. This may be expected if you are loading a checkpoint that shares a subset of the architecture (e.g., loading a `sam2_video` checkpoint into `Sam2Model`), but is otherwise not supported and can yield errors. Please verify that the checkpoint is compatible with the model you are instantiating.
76
+ [transformers] Unrecognized keys in `rope_parameters` for 'rope_type'='default': {'mrope_interleaved'}
77
+
78
+ [transformers] Qwen2_5_VLForConditionalGeneration LOAD REPORT from: /root/ckpts/UnifiedReward-Think-qwen35-9b
79
+ Key | Status |
80
+ -----------------------------------------------------------------------------------+------------+-----------------------------------------------------------------------------------------------
81
+ model.language_model.layers.{0...30}.linear_attn.dt_bias | UNEXPECTED |
82
+ model.language_model.layers.{0...30}.linear_attn.in_proj_z.weight | UNEXPECTED |
83
+ model.language_model.layers.{0...30}.linear_attn.out_proj.weight | UNEXPECTED |
84
+ model.language_model.layers.{0...30}.linear_attn.in_proj_a.weight | UNEXPECTED |
85
+ model.visual.blocks.{0...26}.norm1.bias | UNEXPECTED |
86
+ model.language_model.layers.{0...30}.linear_attn.norm.weight | UNEXPECTED |
87
+ model.language_model.layers.{0...30}.linear_attn.conv1d.weight | UNEXPECTED |
88
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.q_norm.weight | UNEXPECTED |
89
+ model.visual.blocks.{0...26}.mlp.linear_fc1.bias | UNEXPECTED |
90
+ model.visual.blocks.{0...26}.mlp.linear_fc2.bias | UNEXPECTED |
91
+ model.visual.blocks.{0...26}.mlp.linear_fc2.weight | UNEXPECTED |
92
+ model.visual.merger.linear_fc2.weight | UNEXPECTED |
93
+ model.visual.blocks.{0...26}.norm2.bias | UNEXPECTED |
94
+ model.language_model.layers.{0...30}.linear_attn.in_proj_qkv.weight | UNEXPECTED |
95
+ model.language_model.layers.{0...30}.linear_attn.in_proj_b.weight | UNEXPECTED |
96
+ model.visual.blocks.{0...26}.mlp.linear_fc1.weight | UNEXPECTED |
97
+ model.language_model.layers.{0...30}.linear_attn.A_log | UNEXPECTED |
98
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.k_norm.weight | UNEXPECTED |
99
+ model.visual.merger.norm.bias | UNEXPECTED |
100
+ model.visual.merger.norm.weight | UNEXPECTED |
101
+ model.visual.pos_embed.weight | UNEXPECTED |
102
+ model.visual.patch_embed.proj.bias | UNEXPECTED |
103
+ model.visual.merger.linear_fc1.weight | UNEXPECTED |
104
+ model.visual.merger.linear_fc2.bias | UNEXPECTED |
105
+ model.visual.merger.linear_fc1.bias | UNEXPECTED |
106
+ model.visual.blocks.{0...26}.mlp.down_proj.bias | MISSING |
107
+ model.language_model.layers.{0...31}.self_attn.q_proj.bias | MISSING |
108
+ model.visual.blocks.{0...26}.mlp.gate_proj.bias | MISSING |
109
+ model.language_model.layers.{0...30}.self_attn.o_proj.weight | MISSING |
110
+ model.visual.blocks.{0...26}.mlp.down_proj.weight | MISSING |
111
+ model.visual.merger.mlp.{0, 2}.weight | MISSING |
112
+ model.visual.blocks.{0...26}.mlp.up_proj.weight | MISSING |
113
+ model.language_model.layers.{0...30}.self_attn.q_proj.weight | MISSING |
114
+ model.visual.blocks.{0...26}.mlp.gate_proj.weight | MISSING |
115
+ model.language_model.layers.{0...30}.self_attn.k_proj.weight | MISSING |
116
+ model.visual.blocks.{0...26}.mlp.up_proj.bias | MISSING |
117
+ model.language_model.layers.{0...31}.self_attn.v_proj.bias | MISSING |
118
+ model.language_model.layers.{0...31}.self_attn.k_proj.bias | MISSING |
119
+ model.language_model.layers.{0...30}.self_attn.v_proj.weight | MISSING |
120
+ model.visual.merger.ln_q.weight | MISSING |
121
+ model.visual.merger.mlp.{0, 2}.bias | MISSING |
122
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.q_proj.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([8192, 4096]) vs model:torch.Size([4096, 4096])
123
+
124
+ Notes:
125
+ - UNEXPECTED: can be ignored when loading from different task/architecture; not ok if you expect identical arch.
126
+ - MISSING: those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
127
+ - MISMATCH: ckpt weights were loaded, but they did not match the original empty weight shapes.
128
+ [transformers] You are using a model of type `qwen3_5` to instantiate a model of type `qwen2_vl`. This may be expected if you are loading a checkpoint that shares a subset of the architecture (e.g., loading a `sam2_video` checkpoint into `Sam2Model`), but is otherwise not supported and can yield errors. Please verify that the checkpoint is compatible with the model you are instantiating.
129
+ [transformers] Unrecognized keys in `rope_parameters` for 'rope_type'='default': {'mrope_interleaved'}
130
+
131
+ [transformers] Qwen2VLForConditionalGeneration LOAD REPORT from: /root/ckpts/UnifiedReward-Think-qwen35-9b
132
+ Key | Status |
133
+ -----------------------------------------------------------------------------------+------------+---------------------------------------------------------------------------------------------------------------
134
+ model.language_model.layers.{0...30}.linear_attn.dt_bias | UNEXPECTED |
135
+ model.language_model.layers.{0...30}.linear_attn.in_proj_z.weight | UNEXPECTED |
136
+ model.language_model.layers.{0...30}.linear_attn.out_proj.weight | UNEXPECTED |
137
+ model.language_model.layers.{0...30}.linear_attn.in_proj_a.weight | UNEXPECTED |
138
+ model.language_model.layers.{0...30}.linear_attn.norm.weight | UNEXPECTED |
139
+ model.language_model.layers.{0...30}.linear_attn.conv1d.weight | UNEXPECTED |
140
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.q_norm.weight | UNEXPECTED |
141
+ model.visual.blocks.{0...26}.mlp.linear_fc1.bias | UNEXPECTED |
142
+ model.visual.blocks.{0...26}.mlp.linear_fc2.bias | UNEXPECTED |
143
+ model.visual.blocks.{0...26}.mlp.linear_fc2.weight | UNEXPECTED |
144
+ model.visual.merger.linear_fc2.weight | UNEXPECTED |
145
+ model.language_model.layers.{0...30}.linear_attn.in_proj_qkv.weight | UNEXPECTED |
146
+ model.language_model.layers.{0...30}.linear_attn.in_proj_b.weight | UNEXPECTED |
147
+ model.visual.blocks.{0...26}.mlp.linear_fc1.weight | UNEXPECTED |
148
+ model.language_model.layers.{0...30}.linear_attn.A_log | UNEXPECTED |
149
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.k_norm.weight | UNEXPECTED |
150
+ model.visual.merger.norm.bias | UNEXPECTED |
151
+ model.visual.merger.norm.weight | UNEXPECTED |
152
+ model.visual.pos_embed.weight | UNEXPECTED |
153
+ model.visual.patch_embed.proj.bias | UNEXPECTED |
154
+ model.visual.merger.linear_fc1.weight | UNEXPECTED |
155
+ model.visual.merger.linear_fc2.bias | UNEXPECTED |
156
+ model.visual.merger.linear_fc1.bias | UNEXPECTED |
157
+ model.visual.blocks.{0...26}.mlp.fc2.weight | MISSING |
158
+ model.language_model.layers.{0...31}.self_attn.q_proj.bias | MISSING |
159
+ model.language_model.layers.{0...30}.self_attn.o_proj.weight | MISSING |
160
+ model.visual.blocks.{0...26}.mlp.fc1.weight | MISSING |
161
+ model.visual.merger.mlp.{0, 2}.weight | MISSING |
162
+ model.language_model.layers.{0...30}.self_attn.q_proj.weight | MISSING |
163
+ model.language_model.layers.{0...30}.self_attn.k_proj.weight | MISSING |
164
+ model.language_model.layers.{0...31}.self_attn.v_proj.bias | MISSING |
165
+ model.visual.blocks.{0...26}.mlp.fc1.bias | MISSING |
166
+ model.language_model.layers.{0...31}.self_attn.k_proj.bias | MISSING |
167
+ model.visual.blocks.{0...26}.mlp.fc2.bias | MISSING |
168
+ model.language_model.layers.{0...30}.self_attn.v_proj.weight | MISSING |
169
+ model.visual.merger.ln_q.weight | MISSING |
170
+ model.visual.merger.ln_q.bias | MISSING |
171
+ model.visual.merger.mlp.{0, 2}.bias | MISSING |
172
+ model.visual.blocks.{0...26}.attn.qkv.bias | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([3456]) vs model:torch.Size([3840])
173
+ model.visual.blocks.{0...26}.attn.proj.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152, 1152]) vs model:torch.Size([1280, 1280])
174
+ model.visual.blocks.{0...26}.norm1.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152]) vs model:torch.Size([1280])
175
+ model.visual.blocks.{0...26}.norm1.bias | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152]) vs model:torch.Size([1280])
176
+ model.language_model.layers.{3, 7, 11, 15, 19, 23, 27, 31}.self_attn.q_proj.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([8192, 4096]) vs model:torch.Size([4096, 4096])
177
+ model.visual.blocks.{0...26}.attn.proj.bias | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152]) vs model:torch.Size([1280])
178
+ model.visual.blocks.{0...26}.norm2.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152]) vs model:torch.Size([1280])
179
+ model.visual.blocks.{0...26}.norm2.bias | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152]) vs model:torch.Size([1280])
180
+ model.visual.blocks.{0...26}.attn.qkv.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([3456, 1152]) vs model:torch.Size([3840, 1280])
181
+ model.visual.patch_embed.proj.weight | MISMATCH | Reinit due to size mismatch - ckpt: torch.Size([1152, 3, 2, 16, 16]) vs model:torch.Size([1280, 3, 2, 16, 16])
182
+
183
+ Notes:
184
+ - UNEXPECTED: can be ignored when loading from different task/architecture; not ok if you expect identical arch.
185
+ - MISSING: those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
186
+ - MISMATCH: ckpt weights were loaded, but they did not match the original empty weight shapes.
187
+ [transformers] The fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and https://github.com/Dao-AILab/causal-conv1d
188
+
189
+
190
+ [2026-05-21 19:29:33] DIR_B: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100
191
+ [2026-05-21 19:29:33] MODEL_A_NAME: yw002-1-sage-baseline
192
+ [2026-05-21 19:29:33] MODEL_B_NAME: yw006-4-sage-qwen35-9b-token-credit-video-adv-2
193
+ [2026-05-21 19:29:33] OUTPUT_DIR: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100
194
+ [2026-05-21 19:29:33] REWARD_MODEL_PATH: /root/ckpts/UnifiedReward-Think-qwen35-9b
195
+ [2026-05-21 19:29:33] INSTALL_DEPS: true
196
+ [2026-05-21 19:29:33] INSTALL_TORCH: auto
197
+ [2026-05-21 19:29:33] DEVICE_MAP: auto
198
+ [2026-05-21 19:29:33] torch is already available.
199
+ [2026-05-21 19:29:33] Python client/model deps are already available.
200
+ dir_a videos: 300
201
+ dir_b videos: 300
202
+ paired videos selected for this shard: 300
203
+ already completed: 17
204
+ remaining pairs: 283
205
+ reward_model_path: /root/ckpts/UnifiedReward-Think-qwen35-9b
206
+ Detected reward checkpoint: model_type=qwen3_5, architectures=['Qwen3_5ForConditionalGeneration']
207
+ Candidate model classes: Qwen3_5ForConditionalGeneration, AutoModelForImageTextToText
208
+ [transformers] The fast path is not available because one of the required library is not installed. Falling back to torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and https://github.com/Dao-AILab/causal-conv1d
209
+
210
+
211
+ {
212
+ "total": 300,
213
+ "success": 300,
214
+ "counts": {
215
+ "yw002-1-sage-baseline": 183,
216
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 117,
217
+ "tie": 0,
218
+ "unknown": 0,
219
+ "failed": 0
220
+ },
221
+ "win_rate_including_ties": {
222
+ "yw002-1-sage-baseline": 0.61,
223
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 0.39
224
+ },
225
+ "win_rate_decisive_only": {
226
+ "yw002-1-sage-baseline": 0.61,
227
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 0.39
228
+ },
229
+ "latency_sec": {
230
+ "mean": 39.19051666666667,
231
+ "median": 38.563,
232
+ "min": 1.875,
233
+ "max": 84.237
234
+ },
235
+ "dir_a": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100",
236
+ "dir_b": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100",
237
+ "reward_model_path": "/root/ckpts/UnifiedReward-Think-qwen35-9b",
238
+ "output_jsonl": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/pairwise_results.jsonl",
239
+ "shard_count": 1,
240
+ "shard_index": 0
241
+ }
242
+ [2026-05-21 22:35:10] Done.
243
+ [2026-05-21 22:35:10] Results: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/pairwise_results.jsonl
244
+ [2026-05-21 22:35:10] Summary: /nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/summary.json
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/pairwise_results.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/summary.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "total": 300,
3
+ "success": 300,
4
+ "counts": {
5
+ "yw002-1-sage-baseline": 183,
6
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 117,
7
+ "tie": 0,
8
+ "unknown": 0,
9
+ "failed": 0
10
+ },
11
+ "win_rate_including_ties": {
12
+ "yw002-1-sage-baseline": 0.61,
13
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 0.39
14
+ },
15
+ "win_rate_decisive_only": {
16
+ "yw002-1-sage-baseline": 0.61,
17
+ "yw006-4-sage-qwen35-9b-token-credit-video-adv-2": 0.39
18
+ },
19
+ "latency_sec": {
20
+ "mean": 39.19051666666667,
21
+ "median": 38.563,
22
+ "min": 1.875,
23
+ "max": 84.237
24
+ },
25
+ "dir_a": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw002-1-sage-baseline/videogen_val/step_000100",
26
+ "dir_b": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-2/videogen_val/step_000100",
27
+ "reward_model_path": "/root/ckpts/UnifiedReward-Think-qwen35-9b",
28
+ "output_jsonl": "/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/unifiedreward_think_pairwise/yw002-1-sage-baseline_vs_yw006-4-sage-qwen35-9b-token-credit-video-adv-2/step_000100/pairwise_results.jsonl",
29
+ "shard_count": 1,
30
+ "shard_index": 0
31
+ }