Add files using upload-large-folder tool

LTA_openwebtext_dualt/logs/elf_lm1b_t5small_elfb_len128_4gpu_smoke20_20260513.log

@@ -0,0 +1,34 @@
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+You are using the default legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This is expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you. If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it means, and thoroughly read the reason why this was added as explained in https://github.com/huggingface/transformers/pull/24565
+[elf-lm1b] encoder=/e2e-data/evad-tech-vla/wanghan58/models/hf/t5-small enc_dim=512 vocab=32100
+[elf-lm1b] batch=16 world=4 grad_accum=8 gbs~=512
+/usr/local/lib/python3.12/dist-packages/apex/_autocast_utils.py:26: FutureWarning: `torch.cuda.amp.autocast_mode._cast(value, dtype)` is deprecated. Please use `torch.amp.autocast_mode._cast(value, 'cuda', dtype)` instead.
+  return torch.cuda.amp.autocast_mode._cast(args, torch.get_autocast_gpu_dtype())
+/usr/local/lib/python3.12/dist-packages/apex/_autocast_utils.py:26: FutureWarning: `torch.cuda.amp.autocast_mode._cast(value, dtype)` is deprecated. Please use `torch.amp.autocast_mode._cast(value, 'cuda', dtype)` instead.
+  return torch.cuda.amp.autocast_mode._cast(args, torch.get_autocast_gpu_dtype())
+/usr/local/lib/python3.12/dist-packages/apex/_autocast_utils.py:26: FutureWarning: `torch.cuda.amp.autocast_mode._cast(value, dtype)` is deprecated. Please use `torch.amp.autocast_mode._cast(value, 'cuda', dtype)` instead.
+  return torch.cuda.amp.autocast_mode._cast(args, torch.get_autocast_gpu_dtype())
+/usr/local/lib/python3.12/dist-packages/apex/_autocast_utils.py:26: FutureWarning: `torch.cuda.amp.autocast_mode._cast(value, dtype)` is deprecated. Please use `torch.amp.autocast_mode._cast(value, 'cuda', dtype)` instead.
+  return torch.cuda.amp.autocast_mode._cast(args, torch.get_autocast_gpu_dtype())
+[2026-05-13 16:31:30] step=1 elapsed=1.2s lr=4.000000e-04 loss=3.0106 l2=1.7109 ce=1.2996 decoder_frac=0.125 t=0.314 tokens=595
+[2026-05-13 16:31:30] step=2 elapsed=0.4s lr=6.000000e-04 loss=2.7744 l2=1.4773 ce=1.2972 decoder_frac=0.125 t=0.284 tokens=554
+[2026-05-13 16:31:31] step=3 elapsed=0.4s lr=8.000000e-04 loss=3.8823 l2=1.2978 ce=2.5844 decoder_frac=0.250 t=0.389 tokens=561
+[2026-05-13 16:31:31] step=4 elapsed=0.4s lr=1.000000e-03 loss=2.8143 l2=1.5264 ce=1.2878 decoder_frac=0.125 t=0.295 tokens=565
+[2026-05-13 16:31:32] step=5 elapsed=0.4s lr=1.000000e-03 loss=3.7896 l2=1.2306 ce=2.5590 decoder_frac=0.250 t=0.381 tokens=530
+[2026-05-13 16:31:32] step=6 elapsed=0.4s lr=1.000000e-03 loss=2.7638 l2=1.4906 ce=1.2732 decoder_frac=0.125 t=0.303 tokens=525
+[2026-05-13 16:31:33] step=7 elapsed=0.4s lr=1.000000e-03 loss=1.7470 l2=1.7470 ce=0.0000 decoder_frac=0.000 t=0.211 tokens=550
+[2026-05-13 16:31:33] step=8 elapsed=0.4s lr=1.000000e-03 loss=3.9152 l2=1.3883 ce=2.5269 decoder_frac=0.250 t=0.412 tokens=560
+[2026-05-13 16:31:33] step=9 elapsed=0.4s lr=1.000000e-03 loss=3.7428 l2=1.2287 ce=2.5141 decoder_frac=0.250 t=0.395 tokens=576
+[2026-05-13 16:31:34] step=10 elapsed=0.4s lr=1.000000e-03 loss=2.8288 l2=1.5745 ce=1.2543 decoder_frac=0.125 t=0.315 tokens=532
+[2026-05-13 16:31:34] step=11 elapsed=0.4s lr=1.000000e-03 loss=3.7346 l2=1.2263 ce=2.5083 decoder_frac=0.250 t=0.401 tokens=585
+[2026-05-13 16:31:35] step=12 elapsed=0.4s lr=1.000000e-03 loss=3.8412 l2=1.3451 ce=2.4961 decoder_frac=0.250 t=0.407 tokens=550
+[2026-05-13 16:31:35] step=13 elapsed=0.4s lr=1.000000e-03 loss=3.8036 l2=1.3080 ce=2.4955 decoder_frac=0.250 t=0.409 tokens=590
+[2026-05-13 16:31:35] step=14 elapsed=0.4s lr=1.000000e-03 loss=2.7358 l2=1.4966 ce=1.2392 decoder_frac=0.125 t=0.302 tokens=540
+[2026-05-13 16:31:36] step=15 elapsed=0.4s lr=1.000000e-03 loss=4.7200 l2=1.0126 ce=3.7075 decoder_frac=0.375 t=0.494 tokens=529
+[2026-05-13 16:31:36] step=16 elapsed=0.4s lr=1.000000e-03 loss=3.7269 l2=1.2624 ce=2.4645 decoder_frac=0.250 t=0.408 tokens=569
+[2026-05-13 16:31:37] step=17 elapsed=0.4s lr=1.000000e-03 loss=2.7924 l2=1.5645 ce=1.2279 decoder_frac=0.125 t=0.313 tokens=557
+[2026-05-13 16:31:37] step=18 elapsed=0.4s lr=1.000000e-03 loss=2.7998 l2=1.5674 ce=1.2323 decoder_frac=0.125 t=0.324 tokens=540
+[2026-05-13 16:31:38] step=19 elapsed=0.4s lr=1.000000e-03 loss=1.8059 l2=1.8059 ce=0.0000 decoder_frac=0.000 t=0.231 tokens=541
+[2026-05-13 16:31:38] step=20 elapsed=0.4s lr=1.000000e-03 loss=3.7640 l2=1.3074 ce=2.4566 decoder_frac=0.250 t=0.416 tokens=569

LTA_openwebtext_dualt/logs/infer_owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large_resume_20260520_201156.log

@@ -0,0 +1,87 @@
+[sweep] run=runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817
+[skip] step=20359 existing=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step20359_steps128_c1024_t1p45.jsonl
+[infer] step=40718 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step40718_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0040718.pt step=40718
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0040718.pt", "step": 40718, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 1.002838985729012, "nll_per_token": 0.0028349634200807603, "tokens": 504, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 1.002838985729012, "nll_per_token": 0.0028349634200807603, "tokens": 504, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 0.0, "unique_tokens": 1, "token_count": 8192, "distinct_1": 0.0001220703125, "distinct_2": 0.00012218963831867058, "top_token_mass": 1.0}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step40718_steps128_c1024_t1p45.jsonl
+[infer] step=61077 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step61077_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0061077.pt step=61077
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0061077.pt", "step": 61077, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 105.9955547549316, "nll_per_token": 4.663397156958487, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 105.9955547549316, "nll_per_token": 4.663397156958487, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 2.40693613511493, "unique_tokens": 143, "token_count": 8192, "distinct_1": 0.0174560546875, "distinct_2": 0.10117302052785923, "top_token_mass": 0.2647705078125}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step61077_steps128_c1024_t1p45.jsonl
+[infer] step=81436 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step81436_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0081436.pt step=81436
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0081436.pt", "step": 81436, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 20.025109430324235, "nll_per_token": 2.9969869576248467, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 20.02894006760397, "nll_per_token": 2.9971782310336246, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 1.6417594596790073, "unique_tokens": 63, "token_count": 8192, "distinct_1": 0.0076904296875, "distinct_2": 0.056329423264907134, "top_token_mass": 0.4775390625}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step81436_steps128_c1024_t1p45.jsonl
+[infer] step=101795 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step101795_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0101795.pt step=101795
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0101795.pt", "step": 101795, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 13.46160283327985, "nll_per_token": 2.599841398351333, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 6.77012163217799, "nll_per_token": 1.9125190531110905, "tokens": 2022, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 0.7844788420542315, "unique_tokens": 9, "token_count": 8192, "distinct_1": 0.0010986328125, "distinct_2": 0.0039100684261974585, "top_token_mass": 0.6846923828125}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step101795_steps128_c1024_t1p45.jsonl
+[infer] step=122154 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step122154_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0122154.pt step=122154
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0122154.pt", "step": 122154, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 1.0, "nll_per_token": 0.0, "tokens": 0, "kept_samples": 0, "total_samples": 0, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 1.0, "nll_per_token": 0.0, "tokens": 0, "kept_samples": 0, "total_samples": 0, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 0.0, "unique_tokens": 1, "token_count": 8192, "distinct_1": 0.0001220703125, "distinct_2": 0.00012218963831867058, "top_token_mass": 1.0}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step122154_steps128_c1024_t1p45.jsonl
+[infer] step=142513 out=docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step142513_steps128_c1024_t1p45.jsonl
+[ckpt] runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0142513.pt step=142513
+[decode-base] n=8 max_len=1024 steps=128 model_t=post
+[decode-time] schedule=linear s=[0.0,0.25] gumbel=(2.2,0.8) force_final=True t0=0.000000 t_mid=0.500000 t_end=1.000000 dt_mean=0.007812 dt_max=0.007812
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 1/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 2/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 3/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 4/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 5/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 6/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 7/8
+[decode] temp=1.45 final=state rule=dual_line_resample support=1 semantic=1 anchor=state cfg=0/1@0:uniform decode_freq_penalty=0/0/0-1^1 final_sample=argmax/1/k64/p0.95 freq_penalty=0/0/0 start_t=0 start_init=noise time_path=0.0000->1.0000 generated 8/8
+[summary] {"type": "summary", "checkpoint": "runs/lta_owt_compact_gpt2bpe_v8192_stream1024_fullycoupled_rmsnorm_nobias_adamw_wd0p1_uniformt_hardce_mask0p1-1p0_fp32_ddit768x12_gbs512_8gpu_1m_20260519_201817/step_0142513.pt", "step": 142513, "decode": {"steps": 128, "model_t_mode": "post", "decode_time_schedule": "linear", "decode_s_min_frac": 0.0, "decode_s_max_frac": 0.25, "decode_force_final_t": true, "decode_time_grid": [0.0, 0.0078125, 0.015625, 0.0234375, 0.03125, 0.0390625, 0.046875, 0.0546875, 0.0625, 0.0703125, 0.078125, 0.0859375, 0.09375, 0.1015625, 0.109375, 0.1171875, 0.125, 0.1328125, 0.140625, 0.1484375, 0.15625, 0.1640625, 0.171875, 0.1796875, 0.1875, 0.1953125, 0.203125, 0.2109375, 0.21875, 0.2265625, 0.234375, 0.2421875, 0.25, 0.2578125, 0.265625, 0.2734375, 0.28125, 0.2890625, 0.296875, 0.3046875, 0.3125, 0.3203125, 0.328125, 0.3359375, 0.34375, 0.3515625, 0.359375, 0.3671875, 0.375, 0.3828125, 0.390625, 0.3984375, 0.40625, 0.4140625, 0.421875, 0.4296875, 0.4375, 0.4453125, 0.453125, 0.4609375, 0.46875, 0.4765625, 0.484375, 0.4921875, 0.5, 0.5078125, 0.515625, 0.5234375, 0.53125, 0.5390625, 0.546875, 0.5546875, 0.5625, 0.5703125, 0.578125, 0.5859375, 0.59375, 0.6015625, 0.609375, 0.6171875, 0.625, 0.6328125, 0.640625, 0.6484375, 0.65625, 0.6640625, 0.671875, 0.6796875, 0.6875, 0.6953125, 0.703125, 0.7109375, 0.71875, 0.7265625, 0.734375, 0.7421875, 0.75, 0.7578125, 0.765625, 0.7734375, 0.78125, 0.7890625, 0.796875, 0.8046875, 0.8125, 0.8203125, 0.828125, 0.8359375, 0.84375, 0.8515625, 0.859375, 0.8671875, 0.875, 0.8828125, 0.890625, 0.8984375, 0.90625, 0.9140625, 0.921875, 0.9296875, 0.9375, 0.9453125, 0.953125, 0.9609375, 0.96875, 0.9765625, 0.984375, 0.9921875, 1.0], "decode_rule": "dual_line_resample", "support_power": 1.0, "semantic_power": 1.0, "anchor_mode": "state", "cfg_scale": 0.0, "cfg_power": 1.0, "cfg_start": 0.0, "cfg_prior": "uniform", "decode_freq_penalty_alpha": 0.0, "decode_freq_penalty_beta": 0.0, "decode_freq_penalty_floor": 0.0, "decode_freq_penalty_start": 0.0, "decode_freq_penalty_end": 1.0, "decode_freq_penalty_power": 1.0, "start_t": 0.0, "start_init": "noise", "noise_init": "dirichlet", "noise_sigma": -1.0, "dirichlet_concentration": 1.0, "concentration_min": 1.0, "concentration_max": 1024.0, "target_prob": 1.0, "endpoint_temp": 1.45, "final_from": "state", "final_sample_mode": "argmax", "final_sample_temp": 1.0, "final_top_k": 64, "final_top_p": 0.95, "final_freq_penalty_alpha": 0.0, "final_freq_penalty_beta": 0.0, "final_freq_penalty_floor": 0.0, "lock_bos": false, "n_samples": 8, "seed": 20260520}, "raw_genppl": {"ppl": 5.245929944176233, "nll_per_token": 1.657452527214499, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "stripped_genppl": {"ppl": 5.293541278301558, "nll_per_token": 1.666487450693168, "tokens": 2040, "kept_samples": 8, "total_samples": 8, "empty_rate": 0.0, "skipped_samples": 0}, "diversity": {"sample_entropy": 1.2322328503813487, "unique_tokens": 55, "token_count": 8192, "distinct_1": 0.0067138671875, "distinct_2": 0.02761485826001955, "top_token_mass": 0.4080810546875}}
+[done] docs/lta_samples/metrics_20260520/owt_compact_v8192_ckpt_sweep_steps128_c1024_t1p45_n8_large/step142513_steps128_c1024_t1p45.jsonl
+step	raw	stripped	entropy	unique	top_mass

LTA_openwebtext_dualt/logs/lta_owt_gpt2cached_len1024_ddit768x12_elfopt_only_muon_ema_gbs512_8gpu_1m_20260513_032325.log

@@ -0,0 +1,127 @@
+{
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 1,
+  "samples": "wrapped_stream",
+  "vocab_size": 50257,
+  "tokenizer_vocab_size": 50257,
+  "save_dir": "runs/lta_owt_gpt2cached_len1024_ddit768x12_elfopt_only_muon_ema_gbs512_8gpu_1m_20260513_032325",
+  "batch_size": 32,
+  "grad_accum": 16,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "lr_schedule": "constant_warmup",
+  "optimizer": "adamw",
+  "warmup_steps": 2000,
+  "min_lr": 6e-05,
+  "weight_decay": 0.0,
+  "adamw_param_groups": "nanogpt",
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "adam_eps": 1e-08,
+  "muon_momentum": 0.95,
+  "muon_ns_steps": 5,
+  "muon_update_scale": 1.0,
+  "ema_decay": 0.0,
+  "ema_start_step": 0,
+  "model_type": "ddit",
+  "dual_t": true,
+  "corrupt_t_mode": "same",
+  "corrupt_min_t": 0.0,
+  "corrupt_max_t": 1.0,
+  "prefix_block_prob": 0.0,
+  "prefix_block_len": 128,
+  "dirichlet_endpoint_mode": "categorical_dual_t",
+  "dirichlet_semantic_t_mode": "same",
+  "dirichlet_semantic_t_value": 0.0,
+  "categorical_wrong_from_full_vocab": true,
+  "categorical_wrong_from_batch_valid_tokens": false,
+  "mask_mixture_original_prob": 0.0,
+  "mask_mixture_lowk_prob": 0.0,
+  "mask_mixture_lowcorrupt_prob": 0.0,
+  "mask_mixture_block_prob": 0.0,
+  "mask_mixture_all_prob": 0.0,
+  "mask_mixture_lowk_clean_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_lowcorrupt_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_block_tokens": "64,128",
+  "simplex_bridge_sampler": "dirichlet",
+  "logistic_normal_sigma_min": 0.18,
+  "logistic_normal_sigma_max": 2.2,
+  "logistic_normal_tau_min": 0.65,
+  "logistic_normal_tau_max": 1.15,
+  "torch_compile": false,
+  "compile_mode": "max-autotune",
+  "state_format": "prob",
+  "target_loss": "hard_ce",
+  "meanflow_weight": 0.0,
+  "bridge_noise_init": "logistic_normal",
+  "noise_sigma": -1.0,
+  "allow_tf32": true,
+  "activation_checkpointing": false,
+  "activation_checkpoint_interval": 1,
+  "ddp_static_graph": false,
+  "ddp_gradient_as_bucket_view": true,
+  "blocking_data_transfer": false,
+  "dataloader_prefetch_factor": 2,
+  "full_train_stats": false,
+  "wrap": true,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "owt_cached_chunks": false,
+  "owt_chunk_cache_dir": "",
+  "owt_chunk_cache_rebuild": false,
+  "owt_chunk_cache_write_batch": 4096,
+  "owt_exact_repeat_per_chunk": 0,
+  "online_chunk_shuffle": false,
+  "online_chunk_shuffle_buffer": 10000,
+  "openwebtext_split": "train_minus_100k",
+  "detokenizer": "auto",
+  "resolved_detokenizer": null,
+  "num_workers": 1,
+  "latest_every": 5000,
+  "resume_path": ""
+}
+Traceback (most recent call last):
+  File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 1235, in <module>
+    main()
+  File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 1114, in main
+    bridge = make_bridge(
+             ^^^^^^^^^^^^
+  File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 619, in make_bridge
+    return make_dirichlet_bridge_batch(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+TypeError: make_dirichlet_bridge_batch() got an unexpected keyword argument 'categorical_wrong_from_batch_valid_tokens'
+E0513 03:23:37.001000 470107 torch/distributed/elastic/multiprocessing/api.py:870] failed (exitcode: 1) local_rank: 0 (pid: 470197) of binary: /usr/bin/python
+Traceback (most recent call last):
+  File "<frozen runpy>", line 198, in _run_module_as_main
+  File "<frozen runpy>", line 88, in _run_code
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 922, in <module>
+    main()
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
+    return f(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 918, in main
+    run(args)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 909, in run
+    elastic_launch(
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
+    return launch_agent(self._config, self._entrypoint, list(args))
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 270, in launch_agent
+    raise ChildFailedError(
+torch.distributed.elastic.multiprocessing.errors.ChildFailedError: 
+============================================================
+train.py FAILED
+------------------------------------------------------------
+Failures:
+  <NO_OTHER_FAILURES>
+------------------------------------------------------------
+Root Cause (first observed failure):
+[0]:
+  time      : 2026-05-13_03:23:37
+  host      : localhost
+  rank      : 0 (local_rank: 0)
+  exitcode  : 1 (pid: 470197)
+  error_file: <N/A>
+  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
+============================================================

LTA_openwebtext_dualt/logs/owt_dirichlet_len1024_Cv_to_2v_gumbel_sde_watch/infer_lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525_step_0001000.log

@@ -0,0 +1,136 @@
+[watch-gumbel] 2026-05-25_16:35:54 infer runs/lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525/step_0001000.pt -> docs/lta_samples/metrics_20260525/owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_every1k_sde_gumbel_topp0.95_tau1.0_to_0.2_blend_c32100_64200_n128/lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525/step_0001000
+[load] runs/lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525/step_0001000.pt
+[ckpt] step=1000
+[sde] generated 2/128
+[sde] generated 4/128
+[sde] generated 6/128
+[sde] generated 8/128
+[sde] generated 10/128
+[sde] generated 12/128
+[sde] generated 14/128
+[sde] generated 16/128
+[sde] generated 18/128
+[sde] generated 20/128
+[sde] generated 22/128
+[sde] generated 24/128
+[sde] generated 26/128
+[sde] generated 28/128
+[sde] generated 30/128
+[sde] generated 32/128
+[sde] generated 34/128
+[sde] generated 36/128
+[sde] generated 38/128
+[sde] generated 40/128
+[sde] generated 42/128
+[sde] generated 44/128
+[sde] generated 46/128
+[sde] generated 48/128
+[sde] generated 50/128
+[sde] generated 52/128
+[sde] generated 54/128
+[sde] generated 56/128
+[sde] generated 58/128
+[sde] generated 60/128
+[sde] generated 62/128
+[sde] generated 64/128
+[sde] generated 66/128
+[sde] generated 68/128
+[sde] generated 70/128
+[sde] generated 72/128
+[sde] generated 74/128
+[sde] generated 76/128
+[sde] generated 78/128
+[sde] generated 80/128
+[sde] generated 82/128
+[sde] generated 84/128
+[sde] generated 86/128
+[sde] generated 88/128
+[sde] generated 90/128
+[sde] generated 92/128
+[sde] generated 94/128
+[sde] generated 96/128
+[sde] generated 98/128
+[sde] generated 100/128
+[sde] generated 102/128
+[sde] generated 104/128
+[sde] generated 106/128
+[sde] generated 108/128
+[sde] generated 110/128
+[sde] generated 112/128
+[sde] generated 114/128
+[sde] generated 116/128
+[sde] generated 118/128
+[sde] generated 120/128
+[sde] generated 122/128
+[sde] generated 124/128
+[sde] generated 126/128
+[sde] generated 128/128
+[score] loading scorer: /e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-large-standard
+[summary] {
+  "type": "summary",
+  "checkpoint": "runs/lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525/step_0001000.pt",
+  "step": 1000,
+  "decode": {
+    "decode_rule": "dirichlet_resample_sde",
+    "steps": 128,
+    "model_t_mode": "support_t",
+    "mean_mode": "endpoint_only",
+    "anchor_gamma": 1.0,
+    "endpoint_floor": 0.0,
+    "concentration_min": 32100.0,
+    "concentration_max": 64200.0,
+    "endpoint_temp": 1.45,
+    "endpoint_temp_start": null,
+    "endpoint_temp_end": null,
+    "endpoint_projection": "gumbel_softmax",
+    "endpoint_top_k": 0,
+    "endpoint_top_p": 0.95,
+    "gumbel_tau_start": 1.0,
+    "gumbel_tau_end": 0.2,
+    "gumbel_noise_scale_start": 1.0,
+    "gumbel_noise_scale_end": 1.0,
+    "ban_special_tokens": false,
+    "banned_endpoint_ids": [],
+    "support_power": 1.0,
+    "semantic_power": 1.0,
+    "noise_init": "dirichlet",
+    "noise_sigma": -1.0,
+    "noise_dirichlet_concentration": 32100.0,
+    "sde_resample": "dirichlet",
+    "logistic_normal_sigma_min": 0.18,
+    "logistic_normal_sigma_max": 3.0,
+    "logistic_normal_tau_min": 0.65,
+    "logistic_normal_tau_max": 1.0,
+    "final_from": "blend_0.5",
+    "n_samples": 128,
+    "seed": 20260524
+  },
+  "raw_genppl": {
+    "ppl": 538.7763431682185,
+    "nll_per_token": 6.2893005370545065,
+    "tokens": 111239,
+    "kept_samples": 128,
+    "total_samples": 128,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "stripped_genppl": {
+    "ppl": 804.1208604588479,
+    "nll_per_token": 6.689749581835244,
+    "tokens": 96338,
+    "kept_samples": 128,
+    "total_samples": 128,
+    "empty_rate": 0.0,
+    "skipped_samples": 0
+  },
+  "diversity": {
+    "sample_entropy": 3.6305307938962628,
+    "unique_tokens": 12697,
+    "token_count": 131072,
+    "distinct_1": 0.09687042236328125,
+    "distinct_2": 0.38274376832844575,
+    "top_token_mass": 0.277618408203125
+  }
+}
+[done] docs/lta_samples/metrics_20260525/owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_every1k_sde_gumbel_topp0.95_tau1.0_to_0.2_blend_c32100_64200_n128/lta_owt_t5elf_dirichlet_len1024_Cv_to_2v_mask1_sameT_gbs512_b32_8gpu_20k_save1k_gumbelwatch_20260525/step_0001000/sde_steps128_samples128_scored.jsonl
+[watch-gumbel] 2026-05-25_16:42:26 done step_0001000

LTA_openwebtext_dualt/logs/owt_dirichlet_len1024_Cv_to_2v_gumbel_sde_watch/processed_lta_owt_t5_absrope_adaln_dirichlet_len1024_Cv_to_2v_mask0p1_1p0_sameT_gbs512_b32_8gpu_1m_save10k_20260525_steps128_c32100_64200_gumbel_t1p45_n128.txt

@@ -0,0 +1,2 @@
+runs/lta_owt_t5_absrope_adaln_dirichlet_len1024_Cv_to_2v_mask0p1_1p0_sameT_gbs512_b32_8gpu_1m_save10k_20260525/step_0010000.pt
+runs/lta_owt_t5_absrope_adaln_dirichlet_len1024_Cv_to_2v_mask0p1_1p0_sameT_gbs512_b32_8gpu_1m_save10k_20260525/step_0020000.pt

LTA_openwebtext_dualt/logs/owt_from_lm1b_c1024_4gpu/lta_owt_distilbert_len1024_init_lm1b1m_posemb_repeat_fully_c1024_adamw_gbs512_4gpu_20k_20260514_120045.log

@@ -0,0 +1,214 @@
+initialized_model_from=runs/lta_lm1b_dirichlet_categorical_fullvocab_c1024_fullycoupled_flmpack_onehot_hardce_ddit_small_len128_gbs512_8gpu_1m_nw0/step_1000000.pt start_step=1
+{
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 4,
+  "samples": "wrapped_stream",
+  "vocab_size": 30522,
+  "tokenizer_vocab_size": 30522,
+  "save_dir": "runs_transfer/lta_owt_distilbert_len1024_init_lm1b1m_posemb_repeat_fully_c1024_adamw_gbs512_4gpu_20k_20260514_120045",
+  "batch_size": 16,
+  "grad_accum": 8,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "lr_schedule": "constant_warmup",
+  "optimizer": "adamw",
+  "warmup_steps": 2500,
+  "min_lr": 6e-05,
+  "weight_decay": 0.0,
+  "adamw_param_groups": "nanogpt",
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "adam_eps": 1e-08,
+  "muon_momentum": 0.95,
+  "muon_ns_steps": 5,
+  "muon_update_scale": 1.0,
+  "ema_decay": 0.0,
+  "ema_start_step": 0,
+  "model_type": "ddit",
+  "dual_t": true,
+  "corrupt_t_mode": "same",
+  "corrupt_min_t": null,
+  "corrupt_max_t": null,
+  "prefix_block_prob": 0.0,
+  "prefix_block_len": 128,
+  "mask_ratio_floor_schedule": "none",
+  "dirichlet_endpoint_mode": "categorical_dual_t",
+  "dirichlet_semantic_t_mode": "same",
+  "dirichlet_semantic_t_value": 0.0,
+  "dirichlet_semantic_t_curve": "linear",
+  "dirichlet_semantic_t_power": 1.0,
+  "endpoint_sequence_random_prob_alpha": 0.0,
+  "categorical_wrong_from_full_vocab": true,
+  "categorical_wrong_from_batch_valid_tokens": false,
+  "mask_mixture_original_prob": 0.0,
+  "mask_mixture_lowk_prob": 0.0,
+  "mask_mixture_lowcorrupt_prob": 0.0,
+  "mask_mixture_block_prob": 0.0,
+  "mask_mixture_all_prob": 0.0,
+  "mask_mixture_lowk_clean_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_lowcorrupt_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_block_tokens": "64,128",
+  "simplex_bridge_sampler": "dirichlet",
+  "logistic_normal_sigma_min": 0.18,
+  "logistic_normal_sigma_max": 2.2,
+  "logistic_normal_tau_min": 0.65,
+  "logistic_normal_tau_max": 1.15,
+  "torch_compile": false,
+  "compile_mode": "max-autotune",
+  "state_format": "prob",
+  "target_loss": "hard_ce",
+  "meanflow_weight": 0.0,
+  "rollout_train_prob": 0.0,
+  "rollout_train_steps": 1,
+  "rollout_train_infer_steps": 64,
+  "rollout_train_temp": 1.45,
+  "rollout_train_max_gamma": 1.0,
+  "rollout_train_corrupt_only": true,
+  "rollout_train_samplewise": false,
+  "rollout_train_compute_always": false,
+  "bridge_noise_init": "logistic_normal",
+  "noise_sigma": -1.0,
+  "allow_tf32": true,
+  "activation_checkpointing": false,
+  "activation_checkpoint_interval": 1,
+  "activation_checkpoint_scope": "block",
+  "ddp_static_graph": false,
+  "ddp_gradient_as_bucket_view": true,
+  "blocking_data_transfer": false,
+  "dataloader_prefetch_factor": 4,
+  "full_train_stats": false,
+  "record_pad_truncate": false,
+  "record_add_eos": false,
+  "record_add_special_tokens": false,
+  "record_pad_token": "pad",
+  "record_shuffle_buffer": 10000,
+  "wrap": true,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "owt_cached_chunks": false,
+  "owt_chunk_cache_dir": "",
+  "owt_chunk_cache_rebuild": false,
+  "owt_chunk_cache_write_batch": 4096,
+  "owt_exact_repeat_per_chunk": 0,
+  "online_chunk_shuffle": false,
+  "online_chunk_shuffle_buffer": 10000,
+  "openwebtext_split": "train_minus_100k",
+  "detokenizer": "auto",
+  "resolved_detokenizer": null,
+  "num_workers": 4,
+  "latest_every": 500,
+  "resume_path": ""
+}
+step=50 micro_steps=400 elapsed=116.6s lr=6.120000e-06 loss=5.3339 loss_recon=5.3339 loss_meanflow=0.0000 mean_model_t=0.4751 mean_corrupt_t=0.4751 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4602 out_g_norm=0.6815 acc_all=0.7375 acc_corrupt=0.4786 corrupt_frac=0.4672 loss_all=2.2783 loss_corrupt=4.1785 acc_corrupt_t_0p0_0p2=0.0980 corrupt_frac_t_0p0_0p2=0.1932 acc_corrupt_t_0p2_0p4=0.3492 corrupt_frac_t_0p2_0p4=0.2118 acc_corrupt_t_0p4_0p6=0.4777 corrupt_frac_t_0p4_0p6=0.2694 acc_corrupt_t_0p6_0p8=0.7390 corrupt_frac_t_0p6_0p8=0.1662 acc_corrupt_t_0p8_1p0=0.8419 corrupt_frac_t_0p8_1p0=0.1595 wrong_frac=0.5099 init_acc_corrupt=0.4611 init_gold_top10=0.4852 init_gold_top100=0.4888
+step=100 micro_steps=800 elapsed=126.7s lr=1.212000e-05 loss=3.8276 loss_recon=3.8276 loss_meanflow=0.0000 mean_model_t=0.5233 mean_corrupt_t=0.5233 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4635 out_g_norm=0.0751 acc_all=0.6972 acc_corrupt=0.5113 corrupt_frac=0.5995 loss_all=2.4146 loss_corrupt=3.7195 acc_corrupt_t_0p0_0p2=0.1217 corrupt_frac_t_0p0_0p2=0.1698 acc_corrupt_t_0p2_0p4=0.3249 corrupt_frac_t_0p2_0p4=0.2843 acc_corrupt_t_0p4_0p6=0.5761 corrupt_frac_t_0p4_0p6=0.1986 acc_corrupt_t_0p6_0p8=0.7320 corrupt_frac_t_0p6_0p8=0.1850 acc_corrupt_t_0p8_1p0=0.9147 corrupt_frac_t_0p8_1p0=0.1623 wrong_frac=0.4992 init_acc_corrupt=0.4556 init_gold_top10=0.4936 init_gold_top100=0.5009
+step=150 micro_steps=1200 elapsed=127.8s lr=1.812000e-05 loss=3.6168 loss_recon=3.6168 loss_meanflow=0.0000 mean_model_t=0.5071 mean_corrupt_t=0.5071 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4640 out_g_norm=0.0317 acc_all=0.7546 acc_corrupt=0.4738 corrupt_frac=0.4382 loss_all=1.8648 loss_corrupt=3.8321 acc_corrupt_t_0p0_0p2=0.1047 corrupt_frac_t_0p0_0p2=0.2834 acc_corrupt_t_0p2_0p4=0.2506 corrupt_frac_t_0p2_0p4=0.1623 acc_corrupt_t_0p4_0p6=0.5184 corrupt_frac_t_0p4_0p6=0.1400 acc_corrupt_t_0p6_0p8=0.7277 corrupt_frac_t_0p6_0p8=0.2260 acc_corrupt_t_0p8_1p0=0.8839 corrupt_frac_t_0p8_1p0=0.1883 wrong_frac=0.5518 init_acc_corrupt=0.4089 init_gold_top10=0.4396 init_gold_top100=0.4465
+step=200 micro_steps=1600 elapsed=128.5s lr=2.412000e-05 loss=3.5417 loss_recon=3.5417 loss_meanflow=0.0000 mean_model_t=0.4927 mean_corrupt_t=0.4927 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4318 out_g_norm=0.0377 acc_all=0.7870 acc_corrupt=0.6161 corrupt_frac=0.5179 loss_all=1.6049 loss_corrupt=2.7910 acc_corrupt_t_0p0_0p2=0.1794 corrupt_frac_t_0p0_0p2=0.1511 acc_corrupt_t_0p2_0p4=0.3535 corrupt_frac_t_0p2_0p4=0.1710 acc_corrupt_t_0p6_0p8=0.7178 corrupt_frac_t_0p6_0p8=0.3984 acc_corrupt_t_0p8_1p0=0.8676 corrupt_frac_t_0p8_1p0=0.2795 wrong_frac=0.4097 init_acc_corrupt=0.5576 init_gold_top10=0.5852 init_gold_top100=0.5900
+step=250 micro_steps=2000 elapsed=128.2s lr=3.012000e-05 loss=3.3704 loss_recon=3.3704 loss_meanflow=0.0000 mean_model_t=0.4893 mean_corrupt_t=0.4893 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.3537 out_g_norm=0.0438 acc_all=0.7093 acc_corrupt=0.5355 corrupt_frac=0.5968 loss_all=2.0977 loss_corrupt=3.2917 acc_corrupt_t_0p0_0p2=0.1426 corrupt_frac_t_0p0_0p2=0.1958 acc_corrupt_t_0p2_0p4=0.3444 corrupt_frac_t_0p2_0p4=0.1321 acc_corrupt_t_0p4_0p6=0.5655 corrupt_frac_t_0p4_0p6=0.3194 acc_corrupt_t_0p6_0p8=0.7412 corrupt_frac_t_0p6_0p8=0.2502 acc_corrupt_t_0p8_1p0=0.9371 corrupt_frac_t_0p8_1p0=0.1025 wrong_frac=0.4984 init_acc_corrupt=0.4681 init_gold_top10=0.4957 init_gold_top100=0.5027
+step=300 micro_steps=2400 elapsed=128.2s lr=3.612000e-05 loss=3.2718 loss_recon=3.2718 loss_meanflow=0.0000 mean_model_t=0.4885 mean_corrupt_t=0.4885 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.2330 out_g_norm=0.0485 acc_all=0.6102 acc_corrupt=0.4058 corrupt_frac=0.6276 loss_all=2.7787 loss_corrupt=4.1649 acc_corrupt_t_0p0_0p2=0.1212 corrupt_frac_t_0p0_0p2=0.3121 acc_corrupt_t_0p2_0p4=0.3580 corrupt_frac_t_0p2_0p4=0.3434 acc_corrupt_t_0p4_0p6=0.5175 corrupt_frac_t_0p4_0p6=0.1254 acc_corrupt_t_0p6_0p8=0.8011 corrupt_frac_t_0p6_0p8=0.1594 acc_corrupt_t_0p8_1p0=0.8780 corrupt_frac_t_0p8_1p0=0.0598 wrong_frac=0.6378 init_acc_corrupt=0.3071 init_gold_top10=0.3533 init_gold_top100=0.3624
+step=350 micro_steps=2800 elapsed=128.9s lr=4.212000e-05 loss=3.0572 loss_recon=3.0572 loss_meanflow=0.0000 mean_model_t=0.5056 mean_corrupt_t=0.5056 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0927 out_g_norm=0.0585 acc_all=0.7524 acc_corrupt=0.5071 corrupt_frac=0.4842 loss_all=1.7742 loss_corrupt=3.4789 acc_corrupt_t_0p0_0p2=0.1040 corrupt_frac_t_0p0_0p2=0.2085 acc_corrupt_t_0p2_0p4=0.3355 corrupt_frac_t_0p2_0p4=0.2972 acc_corrupt_t_0p4_0p6=0.5518 corrupt_frac_t_0p4_0p6=0.1046 acc_corrupt_t_0p6_0p8=0.7827 corrupt_frac_t_0p6_0p8=0.1775 acc_corrupt_t_0p8_1p0=0.8913 corrupt_frac_t_0p8_1p0=0.2122 wrong_frac=0.5311 init_acc_corrupt=0.4258 init_gold_top10=0.4624 init_gold_top100=0.4701
+step=400 micro_steps=3200 elapsed=128.3s lr=4.812000e-05 loss=2.9771 loss_recon=2.9771 loss_meanflow=0.0000 mean_model_t=0.5016 mean_corrupt_t=0.5016 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9699 out_g_norm=0.0609 acc_all=0.6943 acc_corrupt=0.5593 corrupt_frac=0.6827 loss_all=2.1328 loss_corrupt=3.0564 acc_corrupt_t_0p0_0p2=0.1100 corrupt_frac_t_0p0_0p2=0.1984 acc_corrupt_t_0p2_0p4=0.2757 corrupt_frac_t_0p2_0p4=0.2037 acc_corrupt_t_0p4_0p6=0.6867 corrupt_frac_t_0p4_0p6=0.0890 acc_corrupt_t_0p6_0p8=0.7985 corrupt_frac_t_0p6_0p8=0.3500 acc_corrupt_t_0p8_1p0=0.8858 corrupt_frac_t_0p8_1p0=0.1589 wrong_frac=0.4867 init_acc_corrupt=0.4782 init_gold_top10=0.5059 init_gold_top100=0.5117
+step=450 micro_steps=3600 elapsed=128.4s lr=5.412000e-05 loss=2.9486 loss_recon=2.9486 loss_meanflow=0.0000 mean_model_t=0.5026 mean_corrupt_t=0.5026 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9136 out_g_norm=0.0455 acc_all=0.7720 acc_corrupt=0.5810 corrupt_frac=0.5333 loss_all=1.5517 loss_corrupt=2.8319 acc_corrupt_t_0p0_0p2=0.1067 corrupt_frac_t_0p0_0p2=0.2069 acc_corrupt_t_0p2_0p4=0.4462 corrupt_frac_t_0p2_0p4=0.0298 acc_corrupt_t_0p4_0p6=0.6173 corrupt_frac_t_0p4_0p6=0.3867 acc_corrupt_t_0p6_0p8=0.7514 corrupt_frac_t_0p6_0p8=0.2440 acc_corrupt_t_0p8_1p0=0.9318 corrupt_frac_t_0p8_1p0=0.1326 wrong_frac=0.4705 init_acc_corrupt=0.5163 init_gold_top10=0.5245 init_gold_top100=0.5291
+step=500 micro_steps=4000 elapsed=128.5s lr=6.012000e-05 loss=3.0142 loss_recon=3.0142 loss_meanflow=0.0000 mean_model_t=0.4955 mean_corrupt_t=0.4955 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9002 out_g_norm=0.0361 acc_all=0.8072 acc_corrupt=0.6482 corrupt_frac=0.5325 loss_all=1.1912 loss_corrupt=2.1426 acc_corrupt_t_0p0_0p2=0.1981 corrupt_frac_t_0p0_0p2=0.1667 acc_corrupt_t_0p2_0p4=0.4098 corrupt_frac_t_0p2_0p4=0.1664 acc_corrupt_t_0p4_0p6=0.6542 corrupt_frac_t_0p4_0p6=0.0825 acc_corrupt_t_0p6_0p8=0.7455 corrupt_frac_t_0p6_0p8=0.2396 acc_corrupt_t_0p8_1p0=0.9119 corrupt_frac_t_0p8_1p0=0.3448 wrong_frac=0.4191 init_acc_corrupt=0.5616 init_gold_top10=0.5769 init_gold_top100=0.5794
+step=550 micro_steps=4400 elapsed=132.2s lr=6.612000e-05 loss=2.7808 loss_recon=2.7808 loss_meanflow=0.0000 mean_model_t=0.5161 mean_corrupt_t=0.5161 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.8984 out_g_norm=0.0314 acc_all=0.7191 acc_corrupt=0.4947 corrupt_frac=0.5416 loss_all=1.9256 loss_corrupt=3.4242 acc_corrupt_t_0p0_0p2=0.1651 corrupt_frac_t_0p0_0p2=0.2641 acc_corrupt_t_0p2_0p4=0.3936 corrupt_frac_t_0p2_0p4=0.2055 acc_corrupt_t_0p4_0p6=0.5540 corrupt_frac_t_0p4_0p6=0.2130 acc_corrupt_t_0p6_0p8=0.7080 corrupt_frac_t_0p6_0p8=0.2037 acc_corrupt_t_0p8_1p0=0.9504 corrupt_frac_t_0p8_1p0=0.1136 wrong_frac=0.5721 init_acc_corrupt=0.3973 init_gold_top10=0.4195 init_gold_top100=0.4274
+step=600 micro_steps=4800 elapsed=134.5s lr=7.212000e-05 loss=2.8882 loss_recon=2.8882 loss_meanflow=0.0000 mean_model_t=0.5044 mean_corrupt_t=0.5044 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9031 out_g_norm=0.0293 acc_all=0.7741 acc_corrupt=0.5030 corrupt_frac=0.4395 loss_all=1.5307 loss_corrupt=3.3578 acc_corrupt_t_0p0_0p2=0.1740 corrupt_frac_t_0p0_0p2=0.3209 acc_corrupt_t_0p2_0p4=0.4874 corrupt_frac_t_0p2_0p4=0.1103 acc_corrupt_t_0p4_0p6=0.5979 corrupt_frac_t_0p4_0p6=0.3454 acc_corrupt_t_0p6_0p8=0.8113 corrupt_frac_t_0p6_0p8=0.0912 acc_corrupt_t_0p8_1p0=0.8540 corrupt_frac_t_0p8_1p0=0.1322 wrong_frac=0.5721 init_acc_corrupt=0.3966 init_gold_top10=0.4213 init_gold_top100=0.4265
+step=650 micro_steps=5200 elapsed=135.0s lr=7.812000e-05 loss=2.8381 loss_recon=2.8381 loss_meanflow=0.0000 mean_model_t=0.4989 mean_corrupt_t=0.4989 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9124 out_g_norm=0.0258 acc_all=0.7846 acc_corrupt=0.6305 corrupt_frac=0.5763 loss_all=1.3745 loss_corrupt=2.3498 acc_corrupt_t_0p0_0p2=0.1394 corrupt_frac_t_0p0_0p2=0.1512 acc_corrupt_t_0p2_0p4=0.3558 corrupt_frac_t_0p2_0p4=0.0566 acc_corrupt_t_0p4_0p6=0.5369 corrupt_frac_t_0p4_0p6=0.2681 acc_corrupt_t_0p6_0p8=0.8018 corrupt_frac_t_0p6_0p8=0.3376 acc_corrupt_t_0p8_1p0=0.9364 corrupt_frac_t_0p8_1p0=0.1865 wrong_frac=0.4169 init_acc_corrupt=0.5614 init_gold_top10=0.5769 init_gold_top100=0.5823
+step=700 micro_steps=5600 elapsed=135.4s lr=8.412000e-05 loss=2.8914 loss_recon=2.8914 loss_meanflow=0.0000 mean_model_t=0.5073 mean_corrupt_t=0.5073 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9220 out_g_norm=0.0255 acc_all=0.7479 acc_corrupt=0.5593 corrupt_frac=0.5667 loss_all=1.6396 loss_corrupt=2.8295 acc_corrupt_t_0p0_0p2=0.1840 corrupt_frac_t_0p0_0p2=0.1692 acc_corrupt_t_0p2_0p4=0.4156 corrupt_frac_t_0p2_0p4=0.2674 acc_corrupt_t_0p4_0p6=0.6588 corrupt_frac_t_0p4_0p6=0.3012 acc_corrupt_t_0p6_0p8=0.7360 corrupt_frac_t_0p6_0p8=0.1207 acc_corrupt_t_0p8_1p0=0.9177 corrupt_frac_t_0p8_1p0=0.1414 wrong_frac=0.5222 init_acc_corrupt=0.4458 init_gold_top10=0.4727 init_gold_top100=0.4783
+step=750 micro_steps=6000 elapsed=135.4s lr=9.012000e-05 loss=2.8874 loss_recon=2.8874 loss_meanflow=0.0000 mean_model_t=0.4925 mean_corrupt_t=0.4925 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9276 out_g_norm=0.0243 acc_all=0.7809 acc_corrupt=0.6402 corrupt_frac=0.6071 loss_all=1.4569 loss_corrupt=2.3815 acc_corrupt_t_0p0_0p2=0.3356 corrupt_frac_t_0p0_0p2=0.0440 acc_corrupt_t_0p2_0p4=0.3152 corrupt_frac_t_0p2_0p4=0.1923 acc_corrupt_t_0p4_0p6=0.6658 corrupt_frac_t_0p4_0p6=0.2599 acc_corrupt_t_0p6_0p8=0.7486 corrupt_frac_t_0p6_0p8=0.3908 acc_corrupt_t_0p8_1p0=0.8780 corrupt_frac_t_0p8_1p0=0.1129 wrong_frac=0.4332 init_acc_corrupt=0.5396 init_gold_top10=0.5647 init_gold_top100=0.5691
+step=800 micro_steps=6400 elapsed=135.5s lr=9.612000e-05 loss=2.8735 loss_recon=2.8735 loss_meanflow=0.0000 mean_model_t=0.4870 mean_corrupt_t=0.4870 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9336 out_g_norm=0.0234 acc_all=0.7418 acc_corrupt=0.5373 corrupt_frac=0.5493 loss_all=1.6838 loss_corrupt=2.9801 acc_corrupt_t_0p0_0p2=0.2055 corrupt_frac_t_0p0_0p2=0.2411 acc_corrupt_t_0p2_0p4=0.4128 corrupt_frac_t_0p2_0p4=0.2937 acc_corrupt_t_0p4_0p6=0.6371 corrupt_frac_t_0p4_0p6=0.1001 acc_corrupt_t_0p6_0p8=0.7959 corrupt_frac_t_0p6_0p8=0.3027 acc_corrupt_t_0p8_1p0=0.9911 corrupt_frac_t_0p8_1p0=0.0623 wrong_frac=0.5511 init_acc_corrupt=0.4212 init_gold_top10=0.4455 init_gold_top100=0.4483
+step=850 micro_steps=6800 elapsed=135.2s lr=1.021200e-04 loss=2.7085 loss_recon=2.7085 loss_meanflow=0.0000 mean_model_t=0.5129 mean_corrupt_t=0.5129 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9404 out_g_norm=0.0226 acc_all=0.7911 acc_corrupt=0.6168 corrupt_frac=0.5336 loss_all=1.3531 loss_corrupt=2.4234 acc_corrupt_t_0p0_0p2=0.2197 corrupt_frac_t_0p0_0p2=0.2306 acc_corrupt_t_0p2_0p4=0.3948 corrupt_frac_t_0p2_0p4=0.1283 acc_corrupt_t_0p4_0p6=0.6159 corrupt_frac_t_0p4_0p6=0.2695 acc_corrupt_t_0p6_0p8=0.8349 corrupt_frac_t_0p6_0p8=0.0603 acc_corrupt_t_0p8_1p0=0.9611 corrupt_frac_t_0p8_1p0=0.3113 wrong_frac=0.4631 init_acc_corrupt=0.5093 init_gold_top10=0.5320 init_gold_top100=0.5357
+step=900 micro_steps=7200 elapsed=135.8s lr=1.081200e-04 loss=2.6962 loss_recon=2.6962 loss_meanflow=0.0000 mean_model_t=0.5020 mean_corrupt_t=0.5020 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9474 out_g_norm=0.0237 acc_all=0.8228 acc_corrupt=0.6664 corrupt_frac=0.5297 loss_all=1.1702 loss_corrupt=2.1729 acc_corrupt_t_0p0_0p2=0.1707 corrupt_frac_t_0p0_0p2=0.0722 acc_corrupt_t_0p2_0p4=0.4048 corrupt_frac_t_0p2_0p4=0.1477 acc_corrupt_t_0p4_0p6=0.5952 corrupt_frac_t_0p4_0p6=0.2693 acc_corrupt_t_0p6_0p8=0.8028 corrupt_frac_t_0p6_0p8=0.3506 acc_corrupt_t_0p8_1p0=0.9525 corrupt_frac_t_0p8_1p0=0.1602 wrong_frac=0.4166 init_acc_corrupt=0.5634 init_gold_top10=0.5812 init_gold_top100=0.5843
+step=950 micro_steps=7600 elapsed=135.6s lr=1.141200e-04 loss=2.7332 loss_recon=2.7332 loss_meanflow=0.0000 mean_model_t=0.5095 mean_corrupt_t=0.5095 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9545 out_g_norm=0.0247 acc_all=0.6908 acc_corrupt=0.5107 corrupt_frac=0.6260 loss_all=2.0542 loss_corrupt=3.2336 acc_corrupt_t_0p0_0p2=0.1500 corrupt_frac_t_0p0_0p2=0.1423 acc_corrupt_t_0p2_0p4=0.3361 corrupt_frac_t_0p2_0p4=0.4134 acc_corrupt_t_0p4_0p6=0.5653 corrupt_frac_t_0p4_0p6=0.0500 acc_corrupt_t_0p6_0p8=0.8033 corrupt_frac_t_0p6_0p8=0.3162 acc_corrupt_t_0p8_1p0=0.8727 corrupt_frac_t_0p8_1p0=0.0781 wrong_frac=0.5552 init_acc_corrupt=0.3961 init_gold_top10=0.4404 init_gold_top100=0.4452
+step=1000 micro_steps=8000 elapsed=135.1s lr=1.201200e-04 loss=2.6267 loss_recon=2.6267 loss_meanflow=0.0000 mean_model_t=0.5180 mean_corrupt_t=0.5180 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9606 out_g_norm=0.0223 acc_all=0.8404 acc_corrupt=0.6390 corrupt_frac=0.4393 loss_all=0.9768 loss_corrupt=2.1930 acc_corrupt_t_0p0_0p2=0.0933 corrupt_frac_t_0p0_0p2=0.1326 acc_corrupt_t_0p2_0p4=0.3475 corrupt_frac_t_0p2_0p4=0.1371 acc_corrupt_t_0p4_0p6=0.6511 corrupt_frac_t_0p4_0p6=0.1338 acc_corrupt_t_0p6_0p8=0.7651 corrupt_frac_t_0p6_0p8=0.4016 acc_corrupt_t_0p8_1p0=0.9473 corrupt_frac_t_0p8_1p0=0.1949 wrong_frac=0.4384 init_acc_corrupt=0.5355 init_gold_top10=0.5551 init_gold_top100=0.5627
+step=1050 micro_steps=8400 elapsed=145.5s lr=1.261200e-04 loss=2.7071 loss_recon=2.7071 loss_meanflow=0.0000 mean_model_t=0.5050 mean_corrupt_t=0.5050 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9661 out_g_norm=0.0226 acc_all=0.7328 acc_corrupt=0.5157 corrupt_frac=0.5459 loss_all=1.7703 loss_corrupt=3.1741 acc_corrupt_t_0p0_0p2=0.1688 corrupt_frac_t_0p0_0p2=0.2947 acc_corrupt_t_0p2_0p4=0.3957 corrupt_frac_t_0p2_0p4=0.2461 acc_corrupt_t_0p4_0p6=0.6469 corrupt_frac_t_0p4_0p6=0.1488 acc_corrupt_t_0p6_0p8=0.7754 corrupt_frac_t_0p6_0p8=0.0926 acc_corrupt_t_0p8_1p0=0.9204 corrupt_frac_t_0p8_1p0=0.2178 wrong_frac=0.5662 init_acc_corrupt=0.3931 init_gold_top10=0.4261 init_gold_top100=0.4345
+step=1100 micro_steps=8800 elapsed=162.2s lr=1.321200e-04 loss=2.7524 loss_recon=2.7524 loss_meanflow=0.0000 mean_model_t=0.4982 mean_corrupt_t=0.4982 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9731 out_g_norm=0.0218 acc_all=0.7803 acc_corrupt=0.5817 corrupt_frac=0.5175 loss_all=1.3652 loss_corrupt=2.5602 acc_corrupt_t_0p0_0p2=0.2331 corrupt_frac_t_0p0_0p2=0.1812 acc_corrupt_t_0p2_0p4=0.3940 corrupt_frac_t_0p2_0p4=0.1380 acc_corrupt_t_0p4_0p6=0.5580 corrupt_frac_t_0p4_0p6=0.3418 acc_corrupt_t_0p6_0p8=0.8058 corrupt_frac_t_0p6_0p8=0.1707 acc_corrupt_t_0p8_1p0=0.9320 corrupt_frac_t_0p8_1p0=0.1683 wrong_frac=0.5316 init_acc_corrupt=0.4409 init_gold_top10=0.4631 init_gold_top100=0.4677
+step=1150 micro_steps=9200 elapsed=136.0s lr=1.381200e-04 loss=2.8048 loss_recon=2.8048 loss_meanflow=0.0000 mean_model_t=0.4824 mean_corrupt_t=0.4824 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9701 out_g_norm=0.0228 acc_all=0.7565 acc_corrupt=0.4408 corrupt_frac=0.4284 loss_all=1.5773 loss_corrupt=3.5431 acc_corrupt_t_0p0_0p2=0.2233 corrupt_frac_t_0p0_0p2=0.4390 acc_corrupt_t_0p2_0p4=0.4506 corrupt_frac_t_0p2_0p4=0.2609 acc_corrupt_t_0p4_0p6=0.6832 corrupt_frac_t_0p4_0p6=0.1785 acc_corrupt_t_0p6_0p8=0.7470 corrupt_frac_t_0p6_0p8=0.0710 acc_corrupt_t_0p8_1p0=0.9916 corrupt_frac_t_0p8_1p0=0.0507 wrong_frac=0.6622 init_acc_corrupt=0.2892 init_gold_top10=0.3260 init_gold_top100=0.3357
+step=1200 micro_steps=9600 elapsed=140.0s lr=1.441200e-04 loss=2.6216 loss_recon=2.6216 loss_meanflow=0.0000 mean_model_t=0.4964 mean_corrupt_t=0.4964 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9689 out_g_norm=0.0232 acc_all=0.7330 acc_corrupt=0.6004 corrupt_frac=0.6655 loss_all=1.7570 loss_corrupt=2.6081 acc_corrupt_t_0p0_0p2=0.1446 corrupt_frac_t_0p0_0p2=0.2227 acc_corrupt_t_0p2_0p4=0.4816 corrupt_frac_t_0p2_0p4=0.0897 acc_corrupt_t_0p4_0p6=0.5694 corrupt_frac_t_0p4_0p6=0.2364 acc_corrupt_t_0p6_0p8=0.7657 corrupt_frac_t_0p6_0p8=0.2505 acc_corrupt_t_0p8_1p0=0.9895 corrupt_frac_t_0p8_1p0=0.2007 wrong_frac=0.4577 init_acc_corrupt=0.5181 init_gold_top10=0.5346 init_gold_top100=0.5425
+step=1250 micro_steps=10000 elapsed=135.0s lr=1.501200e-04 loss=2.7294 loss_recon=2.7294 loss_meanflow=0.0000 mean_model_t=0.4896 mean_corrupt_t=0.4896 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9615 out_g_norm=0.0224 acc_all=0.8127 acc_corrupt=0.5698 corrupt_frac=0.4279 loss_all=1.1853 loss_corrupt=2.6737 acc_corrupt_t_0p0_0p2=0.1977 corrupt_frac_t_0p0_0p2=0.2100 acc_corrupt_t_0p2_0p4=0.4490 corrupt_frac_t_0p2_0p4=0.3018 acc_corrupt_t_0p4_0p6=0.5878 corrupt_frac_t_0p4_0p6=0.1471 acc_corrupt_t_0p6_0p8=0.8509 corrupt_frac_t_0p6_0p8=0.1799 acc_corrupt_t_0p8_1p0=0.9505 corrupt_frac_t_0p8_1p0=0.1613 wrong_frac=0.5296 init_acc_corrupt=0.4329 init_gold_top10=0.4628 init_gold_top100=0.4701
+step=1300 micro_steps=10400 elapsed=160.7s lr=1.561200e-04 loss=2.7575 loss_recon=2.7575 loss_meanflow=0.0000 mean_model_t=0.4882 mean_corrupt_t=0.4882 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9627 out_g_norm=0.0216 acc_all=0.6938 acc_corrupt=0.4641 corrupt_frac=0.5604 loss_all=2.0383 loss_corrupt=3.5487 acc_corrupt_t_0p0_0p2=0.1329 corrupt_frac_t_0p0_0p2=0.2442 acc_corrupt_t_0p2_0p4=0.3960 corrupt_frac_t_0p2_0p4=0.3685 acc_corrupt_t_0p4_0p6=0.5903 corrupt_frac_t_0p4_0p6=0.2111 acc_corrupt_t_0p6_0p8=0.8778 corrupt_frac_t_0p6_0p8=0.0838 acc_corrupt_t_0p8_1p0=0.9470 corrupt_frac_t_0p8_1p0=0.0925 wrong_frac=0.6210 init_acc_corrupt=0.3333 init_gold_top10=0.3736 init_gold_top100=0.3785
+step=1350 micro_steps=10800 elapsed=169.7s lr=1.621200e-04 loss=2.6077 loss_recon=2.6077 loss_meanflow=0.0000 mean_model_t=0.5072 mean_corrupt_t=0.5072 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9658 out_g_norm=0.0225 acc_all=0.7194 acc_corrupt=0.5666 corrupt_frac=0.6429 loss_all=1.8358 loss_corrupt=2.8055 acc_corrupt_t_0p0_0p2=0.2255 corrupt_frac_t_0p0_0p2=0.1861 acc_corrupt_t_0p2_0p4=0.2778 corrupt_frac_t_0p2_0p4=0.1476 acc_corrupt_t_0p4_0p6=0.5801 corrupt_frac_t_0p4_0p6=0.2512 acc_corrupt_t_0p6_0p8=0.7721 corrupt_frac_t_0p6_0p8=0.2786 acc_corrupt_t_0p8_1p0=0.8998 corrupt_frac_t_0p8_1p0=0.1364 wrong_frac=0.5107 init_acc_corrupt=0.4538 init_gold_top10=0.4852 init_gold_top100=0.4905
+step=1400 micro_steps=11200 elapsed=145.5s lr=1.681200e-04 loss=2.5496 loss_recon=2.5496 loss_meanflow=0.0000 mean_model_t=0.5090 mean_corrupt_t=0.5090 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9692 out_g_norm=0.0214 acc_all=0.7827 acc_corrupt=0.6595 corrupt_frac=0.6263 loss_all=1.3720 loss_corrupt=2.1213 acc_corrupt_t_0p0_0p2=0.2239 corrupt_frac_t_0p0_0p2=0.2686 acc_corrupt_t_0p2_0p4=0.7283 corrupt_frac_t_0p2_0p4=0.0990 acc_corrupt_t_0p4_0p6=0.6662 corrupt_frac_t_0p4_0p6=0.2333 acc_corrupt_t_0p6_0p8=0.8332 corrupt_frac_t_0p6_0p8=0.1069 acc_corrupt_t_0p8_1p0=0.9676 corrupt_frac_t_0p8_1p0=0.2922 wrong_frac=0.4665 init_acc_corrupt=0.5010 init_gold_top10=0.5258 init_gold_top100=0.5330
+step=1450 micro_steps=11600 elapsed=162.8s lr=1.741200e-04 loss=2.5789 loss_recon=2.5789 loss_meanflow=0.0000 mean_model_t=0.5100 mean_corrupt_t=0.5100 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9557 out_g_norm=0.0212 acc_all=0.8300 acc_corrupt=0.6879 corrupt_frac=0.5385 loss_all=1.0353 loss_corrupt=1.8735 acc_corrupt_t_0p0_0p2=0.3133 corrupt_frac_t_0p0_0p2=0.1418 acc_corrupt_t_0p2_0p4=0.4381 corrupt_frac_t_0p2_0p4=0.1420 acc_corrupt_t_0p4_0p6=0.6334 corrupt_frac_t_0p4_0p6=0.2022 acc_corrupt_t_0p6_0p8=0.7672 corrupt_frac_t_0p6_0p8=0.2128 acc_corrupt_t_0p8_1p0=0.9627 corrupt_frac_t_0p8_1p0=0.3012 wrong_frac=0.4038 init_acc_corrupt=0.5722 init_gold_top10=0.5926 init_gold_top100=0.5958
+step=1500 micro_steps=12000 elapsed=132.3s lr=1.801200e-04 loss=2.6958 loss_recon=2.6958 loss_meanflow=0.0000 mean_model_t=0.4929 mean_corrupt_t=0.4929 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9660 out_g_norm=0.0197 acc_all=0.7805 acc_corrupt=0.5887 corrupt_frac=0.5320 loss_all=1.3908 loss_corrupt=2.5809 acc_corrupt_t_0p0_0p2=0.2995 corrupt_frac_t_0p0_0p2=0.2326 acc_corrupt_t_0p2_0p4=0.4327 corrupt_frac_t_0p2_0p4=0.2277 acc_corrupt_t_0p4_0p6=0.6452 corrupt_frac_t_0p4_0p6=0.0427 acc_corrupt_t_0p6_0p8=0.7559 corrupt_frac_t_0p6_0p8=0.3708 acc_corrupt_t_0p8_1p0=0.8927 corrupt_frac_t_0p8_1p0=0.1262 wrong_frac=0.5092 init_acc_corrupt=0.4466 init_gold_top10=0.4857 init_gold_top100=0.4924
+step=1550 micro_steps=12400 elapsed=127.5s lr=1.861200e-04 loss=2.6536 loss_recon=2.6536 loss_meanflow=0.0000 mean_model_t=0.4986 mean_corrupt_t=0.4986 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9730 out_g_norm=0.0205 acc_all=0.7263 acc_corrupt=0.5079 corrupt_frac=0.5446 loss_all=1.8184 loss_corrupt=3.2197 acc_corrupt_t_0p0_0p2=0.1868 corrupt_frac_t_0p0_0p2=0.3995 acc_corrupt_t_0p2_0p4=0.4309 corrupt_frac_t_0p2_0p4=0.2052 acc_corrupt_t_0p6_0p8=0.7667 corrupt_frac_t_0p6_0p8=0.1609 acc_corrupt_t_0p8_1p0=0.9450 corrupt_frac_t_0p8_1p0=0.2343 wrong_frac=0.5705 init_acc_corrupt=0.3783 init_gold_top10=0.4186 init_gold_top100=0.4284
+step=1600 micro_steps=12800 elapsed=126.0s lr=1.921200e-04 loss=2.5511 loss_recon=2.5511 loss_meanflow=0.0000 mean_model_t=0.4957 mean_corrupt_t=0.4957 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9776 out_g_norm=0.0224 acc_all=0.7740 acc_corrupt=0.5456 corrupt_frac=0.4934 loss_all=1.4688 loss_corrupt=2.9088 acc_corrupt_t_0p0_0p2=0.2934 corrupt_frac_t_0p0_0p2=0.2209 acc_corrupt_t_0p2_0p4=0.3587 corrupt_frac_t_0p2_0p4=0.2425 acc_corrupt_t_0p4_0p6=0.5636 corrupt_frac_t_0p4_0p6=0.2919 acc_corrupt_t_0p8_1p0=0.9373 corrupt_frac_t_0p8_1p0=0.2447 wrong_frac=0.5529 init_acc_corrupt=0.3945 init_gold_top10=0.4406 init_gold_top100=0.4484
+step=1650 micro_steps=13200 elapsed=125.8s lr=1.981200e-04 loss=2.7809 loss_recon=2.7809 loss_meanflow=0.0000 mean_model_t=0.4841 mean_corrupt_t=0.4841 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9821 out_g_norm=0.0201 acc_all=0.8197 acc_corrupt=0.6836 corrupt_frac=0.5569 loss_all=1.1670 loss_corrupt=2.0003 acc_corrupt_t_0p0_0p2=0.2510 corrupt_frac_t_0p0_0p2=0.2161 acc_corrupt_t_0p2_0p4=0.3195 corrupt_frac_t_0p2_0p4=0.0292 acc_corrupt_t_0p4_0p6=0.5628 corrupt_frac_t_0p4_0p6=0.1657 acc_corrupt_t_0p6_0p8=0.7957 corrupt_frac_t_0p6_0p8=0.1727 acc_corrupt_t_0p8_1p0=0.9352 corrupt_frac_t_0p8_1p0=0.4163 wrong_frac=0.3989 init_acc_corrupt=0.5797 init_gold_top10=0.5957 init_gold_top100=0.6002
+step=1700 micro_steps=13600 elapsed=125.9s lr=2.041200e-04 loss=2.5079 loss_recon=2.5079 loss_meanflow=0.0000 mean_model_t=0.5132 mean_corrupt_t=0.5132 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1436.9901 out_g_norm=0.0190 acc_all=0.7779 acc_corrupt=0.6384 corrupt_frac=0.6069 loss_all=1.5086 loss_corrupt=2.4391 acc_corrupt_t_0p0_0p2=0.1793 corrupt_frac_t_0p0_0p2=0.1638 acc_corrupt_t_0p2_0p4=0.3501 corrupt_frac_t_0p2_0p4=0.1896 acc_corrupt_t_0p4_0p6=0.6352 corrupt_frac_t_0p4_0p6=0.0970 acc_corrupt_t_0p6_0p8=0.8059 corrupt_frac_t_0p6_0p8=0.1751 acc_corrupt_t_0p8_1p0=0.9076 corrupt_frac_t_0p8_1p0=0.3745 wrong_frac=0.4287 init_acc_corrupt=0.5396 init_gold_top10=0.5660 init_gold_top100=0.5713
+step=1750 micro_steps=14000 elapsed=126.1s lr=2.101200e-04 loss=2.4587 loss_recon=2.4587 loss_meanflow=0.0000 mean_model_t=0.5126 mean_corrupt_t=0.5126 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0043 out_g_norm=0.0190 acc_all=0.8370 acc_corrupt=0.6146 corrupt_frac=0.4172 loss_all=1.0378 loss_corrupt=2.3970 acc_corrupt_t_0p0_0p2=0.3241 corrupt_frac_t_0p0_0p2=0.1264 acc_corrupt_t_0p2_0p4=0.4424 corrupt_frac_t_0p2_0p4=0.3062 acc_corrupt_t_0p4_0p6=0.5769 corrupt_frac_t_0p4_0p6=0.2054 acc_corrupt_t_0p6_0p8=0.7101 corrupt_frac_t_0p6_0p8=0.0797 acc_corrupt_t_0p8_1p0=0.9321 corrupt_frac_t_0p8_1p0=0.2822 wrong_frac=0.4919 init_acc_corrupt=0.4625 init_gold_top10=0.5033 init_gold_top100=0.5086
+step=1800 micro_steps=14400 elapsed=126.6s lr=2.161200e-04 loss=2.6668 loss_recon=2.6668 loss_meanflow=0.0000 mean_model_t=0.4903 mean_corrupt_t=0.4903 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0161 out_g_norm=0.0189 acc_all=0.8640 acc_corrupt=0.8027 corrupt_frac=0.6876 loss_all=0.8615 loss_corrupt=1.2416 acc_corrupt_t_0p0_0p2=0.3652 corrupt_frac_t_0p0_0p2=0.0316 acc_corrupt_t_0p2_0p4=0.3643 corrupt_frac_t_0p2_0p4=0.1279 acc_corrupt_t_0p6_0p8=0.8018 corrupt_frac_t_0p6_0p8=0.2924 acc_corrupt_t_0p8_1p0=0.9307 corrupt_frac_t_0p8_1p0=0.5481 wrong_frac=0.2614 init_acc_corrupt=0.7203 init_gold_top10=0.7382 init_gold_top100=0.7390
+step=1850 micro_steps=14800 elapsed=126.6s lr=2.221200e-04 loss=2.6053 loss_recon=2.6053 loss_meanflow=0.0000 mean_model_t=0.4987 mean_corrupt_t=0.4987 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0267 out_g_norm=0.0224 acc_all=0.6619 acc_corrupt=0.4466 corrupt_frac=0.5974 loss_all=2.3384 loss_corrupt=3.7637 acc_corrupt_t_0p0_0p2=0.1569 corrupt_frac_t_0p0_0p2=0.4018 acc_corrupt_t_0p2_0p4=0.4129 corrupt_frac_t_0p2_0p4=0.2054 acc_corrupt_t_0p4_0p6=0.5571 corrupt_frac_t_0p4_0p6=0.1190 acc_corrupt_t_0p6_0p8=0.7219 corrupt_frac_t_0p6_0p8=0.1154 acc_corrupt_t_0p8_1p0=0.9413 corrupt_frac_t_0p8_1p0=0.1585 wrong_frac=0.6146 init_acc_corrupt=0.3289 init_gold_top10=0.3739 init_gold_top100=0.3853
+step=1900 micro_steps=15200 elapsed=126.8s lr=2.281200e-04 loss=2.6507 loss_recon=2.6507 loss_meanflow=0.0000 mean_model_t=0.4989 mean_corrupt_t=0.4989 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0399 out_g_norm=0.0205 acc_all=0.6990 acc_corrupt=0.5078 corrupt_frac=0.6080 loss_all=2.1180 loss_corrupt=3.4310 acc_corrupt_t_0p0_0p2=0.1824 corrupt_frac_t_0p0_0p2=0.3347 acc_corrupt_t_0p2_0p4=0.4260 corrupt_frac_t_0p2_0p4=0.1473 acc_corrupt_t_0p4_0p6=0.5835 corrupt_frac_t_0p4_0p6=0.2037 acc_corrupt_t_0p6_0p8=0.7832 corrupt_frac_t_0p6_0p8=0.1792 acc_corrupt_t_0p8_1p0=0.9235 corrupt_frac_t_0p8_1p0=0.1352 wrong_frac=0.5631 init_acc_corrupt=0.3962 init_gold_top10=0.4266 init_gold_top100=0.4361
+step=1950 micro_steps=15600 elapsed=127.1s lr=2.341200e-04 loss=2.6638 loss_recon=2.6638 loss_meanflow=0.0000 mean_model_t=0.4882 mean_corrupt_t=0.4882 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0478 out_g_norm=0.0212 acc_all=0.5979 acc_corrupt=0.4502 corrupt_frac=0.5277 loss_all=2.5492 loss_corrupt=3.5023 acc_corrupt_t_0p0_0p2=0.0795 corrupt_frac_t_0p0_0p2=0.2066 acc_corrupt_t_0p2_0p4=0.1603 corrupt_frac_t_0p2_0p4=0.0967 acc_corrupt_t_0p4_0p6=0.5448 corrupt_frac_t_0p4_0p6=0.3265 acc_corrupt_t_0p6_0p8=0.7777 corrupt_frac_t_0p6_0p8=0.1170 acc_corrupt_t_0p8_1p0=0.5898 corrupt_frac_t_0p8_1p0=0.2532 wrong_frac=0.5185 init_acc_corrupt=0.4648 init_gold_top10=0.4757 init_gold_top100=0.4813
+step=2000 micro_steps=16000 elapsed=127.3s lr=2.401200e-04 loss=2.6270 loss_recon=2.6270 loss_meanflow=0.0000 mean_model_t=0.4961 mean_corrupt_t=0.4961 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0617 out_g_norm=0.0203 acc_all=0.6990 acc_corrupt=0.4510 corrupt_frac=0.5439 loss_all=1.9510 loss_corrupt=3.5173 acc_corrupt_t_0p0_0p2=0.2108 corrupt_frac_t_0p0_0p2=0.2742 acc_corrupt_t_0p2_0p4=0.3763 corrupt_frac_t_0p2_0p4=0.4351 acc_corrupt_t_0p4_0p6=0.6029 corrupt_frac_t_0p4_0p6=0.1085 acc_corrupt_t_0p6_0p8=0.8790 corrupt_frac_t_0p6_0p8=0.0742 acc_corrupt_t_0p8_1p0=0.9148 corrupt_frac_t_0p8_1p0=0.1081 wrong_frac=0.6441 init_acc_corrupt=0.2901 init_gold_top10=0.3463 init_gold_top100=0.3555
+step=2050 micro_steps=16400 elapsed=130.6s lr=2.461200e-04 loss=2.5362 loss_recon=2.5362 loss_meanflow=0.0000 mean_model_t=0.5098 mean_corrupt_t=0.5098 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0771 out_g_norm=0.0224 acc_all=0.7955 acc_corrupt=0.6389 corrupt_frac=0.5554 loss_all=1.2729 loss_corrupt=2.2235 acc_corrupt_t_0p0_0p2=0.1632 corrupt_frac_t_0p0_0p2=0.1582 acc_corrupt_t_0p2_0p4=0.4926 corrupt_frac_t_0p2_0p4=0.2222 acc_corrupt_t_0p4_0p6=0.6393 corrupt_frac_t_0p4_0p6=0.0710 acc_corrupt_t_0p6_0p8=0.8082 corrupt_frac_t_0p6_0p8=0.4400 acc_corrupt_t_0p8_1p0=0.9453 corrupt_frac_t_0p8_1p0=0.1086 wrong_frac=0.4577 init_acc_corrupt=0.5146 init_gold_top10=0.5396 init_gold_top100=0.5435
+step=2100 micro_steps=16800 elapsed=129.7s lr=2.521200e-04 loss=2.5877 loss_recon=2.5877 loss_meanflow=0.0000 mean_model_t=0.5055 mean_corrupt_t=0.5055 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.0957 out_g_norm=0.0188 acc_all=0.7486 acc_corrupt=0.5588 corrupt_frac=0.5657 loss_all=1.5975 loss_corrupt=2.7605 acc_corrupt_t_0p0_0p2=0.1941 corrupt_frac_t_0p0_0p2=0.2763 acc_corrupt_t_0p2_0p4=0.4395 corrupt_frac_t_0p2_0p4=0.2079 acc_corrupt_t_0p4_0p6=0.6538 corrupt_frac_t_0p4_0p6=0.0589 acc_corrupt_t_0p6_0p8=0.7708 corrupt_frac_t_0p6_0p8=0.3441 acc_corrupt_t_0p8_1p0=0.9761 corrupt_frac_t_0p8_1p0=0.1128 wrong_frac=0.5373 init_acc_corrupt=0.4245 init_gold_top10=0.4539 init_gold_top100=0.4622
+step=2150 micro_steps=17200 elapsed=189.8s lr=2.581200e-04 loss=2.6215 loss_recon=2.6215 loss_meanflow=0.0000 mean_model_t=0.4933 mean_corrupt_t=0.4933 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.1193 out_g_norm=0.0182 acc_all=0.8499 acc_corrupt=0.7113 corrupt_frac=0.5118 loss_all=0.9500 loss_corrupt=1.7832 acc_corrupt_t_0p0_0p2=0.3760 corrupt_frac_t_0p0_0p2=0.0625 acc_corrupt_t_0p2_0p4=0.4883 corrupt_frac_t_0p2_0p4=0.0972 acc_corrupt_t_0p4_0p6=0.6377 corrupt_frac_t_0p4_0p6=0.3970 acc_corrupt_t_0p6_0p8=0.8138 corrupt_frac_t_0p6_0p8=0.2056 acc_corrupt_t_0p8_1p0=0.9248 corrupt_frac_t_0p8_1p0=0.2378 wrong_frac=0.3896 init_acc_corrupt=0.6040 init_gold_top10=0.6088 init_gold_top100=0.6113
+step=2200 micro_steps=17600 elapsed=151.9s lr=2.641200e-04 loss=2.5421 loss_recon=2.5421 loss_meanflow=0.0000 mean_model_t=0.5027 mean_corrupt_t=0.5027 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.1430 out_g_norm=0.0190 acc_all=0.7783 acc_corrupt=0.6010 corrupt_frac=0.5510 loss_all=1.4238 loss_corrupt=2.5334 acc_corrupt_t_0p0_0p2=0.1998 corrupt_frac_t_0p0_0p2=0.2146 acc_corrupt_t_0p2_0p4=0.4345 corrupt_frac_t_0p2_0p4=0.1489 acc_corrupt_t_0p4_0p6=0.6114 corrupt_frac_t_0p4_0p6=0.2924 acc_corrupt_t_0p6_0p8=0.8280 corrupt_frac_t_0p6_0p8=0.0689 acc_corrupt_t_0p8_1p0=0.9360 corrupt_frac_t_0p8_1p0=0.2753 wrong_frac=0.4800 init_acc_corrupt=0.4895 init_gold_top10=0.5151 init_gold_top100=0.5214
+step=2250 micro_steps=18000 elapsed=127.3s lr=2.701200e-04 loss=2.4855 loss_recon=2.4855 loss_meanflow=0.0000 mean_model_t=0.5112 mean_corrupt_t=0.5112 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.1690 out_g_norm=0.0175 acc_all=0.8243 acc_corrupt=0.6237 corrupt_frac=0.4653 loss_all=1.0825 loss_corrupt=2.2848 acc_corrupt_t_0p0_0p2=0.2535 corrupt_frac_t_0p0_0p2=0.2312 acc_corrupt_t_0p2_0p4=0.5107 corrupt_frac_t_0p2_0p4=0.2517 acc_corrupt_t_0p4_0p6=0.6739 corrupt_frac_t_0p4_0p6=0.1267 acc_corrupt_t_0p6_0p8=0.8069 corrupt_frac_t_0p6_0p8=0.1284 acc_corrupt_t_0p8_1p0=0.9449 corrupt_frac_t_0p8_1p0=0.2619 wrong_frac=0.4915 init_acc_corrupt=0.4643 init_gold_top10=0.5029 init_gold_top100=0.5097
+step=2300 micro_steps=18400 elapsed=127.3s lr=2.761200e-04 loss=2.6499 loss_recon=2.6499 loss_meanflow=0.0000 mean_model_t=0.4984 mean_corrupt_t=0.4984 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.1925 out_g_norm=0.0173 acc_all=0.7237 acc_corrupt=0.5649 corrupt_frac=0.6260 loss_all=1.7937 loss_corrupt=2.7912 acc_corrupt_t_0p0_0p2=0.1972 corrupt_frac_t_0p0_0p2=0.2481 acc_corrupt_t_0p2_0p4=0.4177 corrupt_frac_t_0p2_0p4=0.3312 acc_corrupt_t_0p4_0p6=0.6667 corrupt_frac_t_0p4_0p6=0.0331 acc_corrupt_t_0p6_0p8=0.7989 corrupt_frac_t_0p6_0p8=0.0892 acc_corrupt_t_0p8_1p0=0.9526 corrupt_frac_t_0p8_1p0=0.2984 wrong_frac=0.5141 init_acc_corrupt=0.4426 init_gold_top10=0.4771 init_gold_top100=0.4852
+step=2350 micro_steps=18800 elapsed=137.3s lr=2.821200e-04 loss=2.6045 loss_recon=2.6045 loss_meanflow=0.0000 mean_model_t=0.5036 mean_corrupt_t=0.5036 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.2200 out_g_norm=0.0197 acc_all=0.8186 acc_corrupt=0.6806 corrupt_frac=0.5667 loss_all=1.1970 loss_corrupt=2.0978 acc_corrupt_t_0p0_0p2=0.1663 corrupt_frac_t_0p0_0p2=0.1081 acc_corrupt_t_0p2_0p4=0.4853 corrupt_frac_t_0p2_0p4=0.1323 acc_corrupt_t_0p4_0p6=0.6358 corrupt_frac_t_0p4_0p6=0.2824 acc_corrupt_t_0p6_0p8=0.8219 corrupt_frac_t_0p6_0p8=0.1917 acc_corrupt_t_0p8_1p0=0.9151 corrupt_frac_t_0p8_1p0=0.2855 wrong_frac=0.4029 init_acc_corrupt=0.5750 init_gold_top10=0.5958 init_gold_top100=0.5980
+step=2400 micro_steps=19200 elapsed=191.9s lr=2.881200e-04 loss=2.5857 loss_recon=2.5857 loss_meanflow=0.0000 mean_model_t=0.5045 mean_corrupt_t=0.5045 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.2521 out_g_norm=0.0172 acc_all=0.8042 acc_corrupt=0.6555 corrupt_frac=0.5606 loss_all=1.3039 loss_corrupt=2.2668 acc_corrupt_t_0p0_0p2=0.2120 corrupt_frac_t_0p0_0p2=0.2157 acc_corrupt_t_0p2_0p4=0.5235 corrupt_frac_t_0p2_0p4=0.1042 acc_corrupt_t_0p4_0p6=0.5674 corrupt_frac_t_0p4_0p6=0.1437 acc_corrupt_t_0p6_0p8=0.8107 corrupt_frac_t_0p6_0p8=0.2703 acc_corrupt_t_0p8_1p0=0.9566 corrupt_frac_t_0p8_1p0=0.2661 wrong_frac=0.4133 init_acc_corrupt=0.5522 init_gold_top10=0.5816 init_gold_top100=0.5875
+step=2450 micro_steps=19600 elapsed=166.9s lr=2.941200e-04 loss=2.6024 loss_recon=2.6024 loss_meanflow=0.0000 mean_model_t=0.4977 mean_corrupt_t=0.4977 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.2734 out_g_norm=0.0186 acc_all=0.8117 acc_corrupt=0.5566 corrupt_frac=0.4149 loss_all=1.2019 loss_corrupt=2.7634 acc_corrupt_t_0p0_0p2=0.2468 corrupt_frac_t_0p0_0p2=0.2319 acc_corrupt_t_0p2_0p4=0.4267 corrupt_frac_t_0p2_0p4=0.1896 acc_corrupt_t_0p4_0p6=0.6373 corrupt_frac_t_0p4_0p6=0.2556 acc_corrupt_t_0p6_0p8=0.7369 corrupt_frac_t_0p6_0p8=0.2500 acc_corrupt_t_0p8_1p0=0.9778 corrupt_frac_t_0p8_1p0=0.0730 wrong_frac=0.5642 init_acc_corrupt=0.3936 init_gold_top10=0.4255 init_gold_top100=0.4355
+step=2500 micro_steps=20000 elapsed=188.7s lr=3.000000e-04 loss=2.5994 loss_recon=2.5994 loss_meanflow=0.0000 mean_model_t=0.4970 mean_corrupt_t=0.4970 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.3060 out_g_norm=0.0186 acc_all=0.7613 acc_corrupt=0.5852 corrupt_frac=0.5687 loss_all=1.5017 loss_corrupt=2.5873 acc_corrupt_t_0p0_0p2=0.1967 corrupt_frac_t_0p0_0p2=0.2565 acc_corrupt_t_0p2_0p4=0.5517 corrupt_frac_t_0p2_0p4=0.0924 acc_corrupt_t_0p4_0p6=0.6318 corrupt_frac_t_0p4_0p6=0.2582 acc_corrupt_t_0p6_0p8=0.8096 corrupt_frac_t_0p6_0p8=0.3782 acc_corrupt_t_0p8_1p0=0.9853 corrupt_frac_t_0p8_1p0=0.0146 wrong_frac=0.5193 init_acc_corrupt=0.4512 init_gold_top10=0.4723 init_gold_top100=0.4807
+step=2550 micro_steps=20400 elapsed=132.6s lr=3.000000e-04 loss=2.5504 loss_recon=2.5504 loss_meanflow=0.0000 mean_model_t=0.4926 mean_corrupt_t=0.4926 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.3398 out_g_norm=0.0175 acc_all=0.8160 acc_corrupt=0.6625 corrupt_frac=0.5410 loss_all=1.1167 loss_corrupt=2.0379 acc_corrupt_t_0p0_0p2=0.0795 corrupt_frac_t_0p0_0p2=0.1022 acc_corrupt_t_0p2_0p4=0.5069 corrupt_frac_t_0p2_0p4=0.0245 acc_corrupt_t_0p4_0p6=0.5825 corrupt_frac_t_0p4_0p6=0.3515 acc_corrupt_t_0p6_0p8=0.8030 corrupt_frac_t_0p6_0p8=0.4049 acc_corrupt_t_0p8_1p0=0.9585 corrupt_frac_t_0p8_1p0=0.1169 wrong_frac=0.4369 init_acc_corrupt=0.5573 init_gold_top10=0.5614 init_gold_top100=0.5633
+step=2600 micro_steps=20800 elapsed=135.9s lr=3.000000e-04 loss=2.6894 loss_recon=2.6894 loss_meanflow=0.0000 mean_model_t=0.4852 mean_corrupt_t=0.4852 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.3696 out_g_norm=0.0182 acc_all=0.7764 acc_corrupt=0.5881 corrupt_frac=0.5316 loss_all=1.4135 loss_corrupt=2.5560 acc_corrupt_t_0p0_0p2=0.2508 corrupt_frac_t_0p0_0p2=0.2665 acc_corrupt_t_0p2_0p4=0.4863 corrupt_frac_t_0p2_0p4=0.1759 acc_corrupt_t_0p4_0p6=0.6473 corrupt_frac_t_0p4_0p6=0.2393 acc_corrupt_t_0p6_0p8=0.8050 corrupt_frac_t_0p6_0p8=0.1643 acc_corrupt_t_0p8_1p0=0.9650 corrupt_frac_t_0p8_1p0=0.1540 wrong_frac=0.5369 init_acc_corrupt=0.4250 init_gold_top10=0.4578 init_gold_top100=0.4631
+step=2650 micro_steps=21200 elapsed=128.1s lr=3.000000e-04 loss=2.4579 loss_recon=2.4579 loss_meanflow=0.0000 mean_model_t=0.5082 mean_corrupt_t=0.5082 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.3994 out_g_norm=0.0159 acc_all=0.7357 acc_corrupt=0.5233 corrupt_frac=0.5531 loss_all=1.7401 loss_corrupt=3.1175 acc_corrupt_t_0p0_0p2=0.1625 corrupt_frac_t_0p0_0p2=0.2499 acc_corrupt_t_0p2_0p4=0.4180 corrupt_frac_t_0p2_0p4=0.1769 acc_corrupt_t_0p4_0p6=0.6472 corrupt_frac_t_0p4_0p6=0.3006 acc_corrupt_t_0p6_0p8=0.7649 corrupt_frac_t_0p6_0p8=0.2347 acc_corrupt_t_0p8_1p0=0.9155 corrupt_frac_t_0p8_1p0=0.0379 wrong_frac=0.5721 init_acc_corrupt=0.3760 init_gold_top10=0.4209 init_gold_top100=0.4289
+step=2700 micro_steps=21600 elapsed=128.0s lr=3.000000e-04 loss=2.5789 loss_recon=2.5789 loss_meanflow=0.0000 mean_model_t=0.5013 mean_corrupt_t=0.5013 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4361 out_g_norm=0.0157 acc_all=0.7776 acc_corrupt=0.6148 corrupt_frac=0.5754 loss_all=1.4402 loss_corrupt=2.4666 acc_corrupt_t_0p0_0p2=0.1722 corrupt_frac_t_0p0_0p2=0.1971 acc_corrupt_t_0p2_0p4=0.5234 corrupt_frac_t_0p2_0p4=0.1019 acc_corrupt_t_0p4_0p6=0.6153 corrupt_frac_t_0p4_0p6=0.3220 acc_corrupt_t_0p6_0p8=0.8310 corrupt_frac_t_0p6_0p8=0.2285 acc_corrupt_t_0p8_1p0=0.9267 corrupt_frac_t_0p8_1p0=0.1505 wrong_frac=0.4891 init_acc_corrupt=0.4898 init_gold_top10=0.5039 init_gold_top100=0.5117
+step=2750 micro_steps=22000 elapsed=175.8s lr=3.000000e-04 loss=2.5256 loss_recon=2.5256 loss_meanflow=0.0000 mean_model_t=0.4917 mean_corrupt_t=0.4917 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4698 out_g_norm=0.0160 acc_all=0.7535 acc_corrupt=0.5414 corrupt_frac=0.5229 loss_all=1.5800 loss_corrupt=2.9021 acc_corrupt_t_0p0_0p2=0.1629 corrupt_frac_t_0p0_0p2=0.2930 acc_corrupt_t_0p2_0p4=0.4371 corrupt_frac_t_0p2_0p4=0.1335 acc_corrupt_t_0p4_0p6=0.6552 corrupt_frac_t_0p4_0p6=0.2658 acc_corrupt_t_0p6_0p8=0.8143 corrupt_frac_t_0p6_0p8=0.2432 acc_corrupt_t_0p8_1p0=0.9783 corrupt_frac_t_0p8_1p0=0.0645 wrong_frac=0.5627 init_acc_corrupt=0.4105 init_gold_top10=0.4299 init_gold_top100=0.4364
+step=2800 micro_steps=22400 elapsed=127.5s lr=3.000000e-04 loss=2.5223 loss_recon=2.5223 loss_meanflow=0.0000 mean_model_t=0.4977 mean_corrupt_t=0.4977 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.4937 out_g_norm=0.0166 acc_all=0.7708 acc_corrupt=0.6060 corrupt_frac=0.5757 loss_all=1.4794 loss_corrupt=2.5187 acc_corrupt_t_0p0_0p2=0.1794 corrupt_frac_t_0p0_0p2=0.2759 acc_corrupt_t_0p2_0p4=0.4118 corrupt_frac_t_0p2_0p4=0.0896 acc_corrupt_t_0p4_0p6=0.6746 corrupt_frac_t_0p4_0p6=0.2362 acc_corrupt_t_0p6_0p8=0.8468 corrupt_frac_t_0p6_0p8=0.1460 acc_corrupt_t_0p8_1p0=0.9378 corrupt_frac_t_0p8_1p0=0.2523 wrong_frac=0.4759 init_acc_corrupt=0.4946 init_gold_top10=0.5192 init_gold_top100=0.5246
+step=2850 micro_steps=22800 elapsed=127.5s lr=3.000000e-04 loss=2.4808 loss_recon=2.4808 loss_meanflow=0.0000 mean_model_t=0.5025 mean_corrupt_t=0.5025 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.5213 out_g_norm=0.0167 acc_all=0.7151 acc_corrupt=0.5529 corrupt_frac=0.6359 loss_all=1.8672 loss_corrupt=2.9033 acc_corrupt_t_0p0_0p2=0.2304 corrupt_frac_t_0p0_0p2=0.1346 acc_corrupt_t_0p2_0p4=0.3629 corrupt_frac_t_0p2_0p4=0.2748 acc_corrupt_t_0p4_0p6=0.5599 corrupt_frac_t_0p4_0p6=0.2340 acc_corrupt_t_0p6_0p8=0.7679 corrupt_frac_t_0p6_0p8=0.1985 acc_corrupt_t_0p8_1p0=0.8774 corrupt_frac_t_0p8_1p0=0.1581 wrong_frac=0.5388 init_acc_corrupt=0.4100 init_gold_top10=0.4572 init_gold_top100=0.4627
+step=2900 micro_steps=23200 elapsed=175.2s lr=3.000000e-04 loss=2.4895 loss_recon=2.4895 loss_meanflow=0.0000 mean_model_t=0.4913 mean_corrupt_t=0.4913 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.5564 out_g_norm=0.0155 acc_all=0.8445 acc_corrupt=0.6641 corrupt_frac=0.4583 loss_all=0.9392 loss_corrupt=1.9931 acc_corrupt_t_0p0_0p2=0.3475 corrupt_frac_t_0p0_0p2=0.1541 acc_corrupt_t_0p2_0p4=0.5428 corrupt_frac_t_0p2_0p4=0.1401 acc_corrupt_t_0p4_0p6=0.6843 corrupt_frac_t_0p4_0p6=0.2016 acc_corrupt_t_0p6_0p8=0.7364 corrupt_frac_t_0p6_0p8=0.3512 acc_corrupt_t_0p8_1p0=0.9017 corrupt_frac_t_0p8_1p0=0.1530 wrong_frac=0.4686 init_acc_corrupt=0.5106 init_gold_top10=0.5282 init_gold_top100=0.5327
+step=2950 micro_steps=23600 elapsed=128.0s lr=3.000000e-04 loss=2.4493 loss_recon=2.4493 loss_meanflow=0.0000 mean_model_t=0.5021 mean_corrupt_t=0.5021 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.5919 out_g_norm=0.0154 acc_all=0.8262 acc_corrupt=0.6643 corrupt_frac=0.5140 loss_all=1.1565 loss_corrupt=2.2020 acc_corrupt_t_0p0_0p2=0.3403 corrupt_frac_t_0p0_0p2=0.0621 acc_corrupt_t_0p2_0p4=0.3602 corrupt_frac_t_0p2_0p4=0.2921 acc_corrupt_t_0p4_0p6=0.6340 corrupt_frac_t_0p4_0p6=0.1891 acc_corrupt_t_0p6_0p8=0.8232 corrupt_frac_t_0p6_0p8=0.2096 acc_corrupt_t_0p8_1p0=0.9938 corrupt_frac_t_0p8_1p0=0.2471 wrong_frac=0.4202 init_acc_corrupt=0.5444 init_gold_top10=0.5761 init_gold_top100=0.5810
+step=3000 micro_steps=24000 elapsed=127.7s lr=3.000000e-04 loss=2.5211 loss_recon=2.5211 loss_meanflow=0.0000 mean_model_t=0.4971 mean_corrupt_t=0.4971 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.6226 out_g_norm=0.0149 acc_all=0.8489 acc_corrupt=0.7070 corrupt_frac=0.5139 loss_all=0.9723 loss_corrupt=1.8654 acc_corrupt_t_0p0_0p2=0.1772 corrupt_frac_t_0p0_0p2=0.1468 acc_corrupt_t_0p2_0p4=0.4983 corrupt_frac_t_0p2_0p4=0.0691 acc_corrupt_t_0p4_0p6=0.6519 corrupt_frac_t_0p4_0p6=0.2010 acc_corrupt_t_0p6_0p8=0.8450 corrupt_frac_t_0p6_0p8=0.3365 acc_corrupt_t_0p8_1p0=0.9374 corrupt_frac_t_0p8_1p0=0.2467 wrong_frac=0.3857 init_acc_corrupt=0.5893 init_gold_top10=0.6100 init_gold_top100=0.6150
+step=3050 micro_steps=24400 elapsed=130.7s lr=3.000000e-04 loss=2.5400 loss_recon=2.5400 loss_meanflow=0.0000 mean_model_t=0.5055 mean_corrupt_t=0.5055 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.6574 out_g_norm=0.0160 acc_all=0.8238 acc_corrupt=0.7101 corrupt_frac=0.6019 loss_all=1.0997 loss_corrupt=1.7888 acc_corrupt_t_0p0_0p2=0.3058 corrupt_frac_t_0p0_0p2=0.0245 acc_corrupt_t_0p2_0p4=0.3496 corrupt_frac_t_0p2_0p4=0.1679 acc_corrupt_t_0p4_0p6=0.6469 corrupt_frac_t_0p4_0p6=0.2430 acc_corrupt_t_0p6_0p8=0.8100 corrupt_frac_t_0p6_0p8=0.3075 acc_corrupt_t_0p8_1p0=0.9243 corrupt_frac_t_0p8_1p0=0.2571 wrong_frac=0.3796 init_acc_corrupt=0.6051 init_gold_top10=0.6190 init_gold_top100=0.6210
+step=3100 micro_steps=24800 elapsed=176.4s lr=3.000000e-04 loss=2.4574 loss_recon=2.4574 loss_meanflow=0.0000 mean_model_t=0.4958 mean_corrupt_t=0.4958 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.6912 out_g_norm=0.0153 acc_all=0.9191 acc_corrupt=0.8415 corrupt_frac=0.4999 loss_all=0.4614 loss_corrupt=0.8822 acc_corrupt_t_0p0_0p2=0.3605 corrupt_frac_t_0p0_0p2=0.0420 acc_corrupt_t_0p2_0p4=0.6080 corrupt_frac_t_0p2_0p4=0.0243 acc_corrupt_t_0p4_0p6=0.7467 corrupt_frac_t_0p4_0p6=0.0554 acc_corrupt_t_0p6_0p8=0.8406 corrupt_frac_t_0p6_0p8=0.5783 acc_corrupt_t_0p8_1p0=0.9471 corrupt_frac_t_0p8_1p0=0.3000 wrong_frac=0.2622 init_acc_corrupt=0.7337 init_gold_top10=0.7370 init_gold_top100=0.7375
+step=3150 micro_steps=25200 elapsed=127.5s lr=3.000000e-04 loss=2.4397 loss_recon=2.4397 loss_meanflow=0.0000 mean_model_t=0.5056 mean_corrupt_t=0.5056 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.7293 out_g_norm=0.0156 acc_all=0.8318 acc_corrupt=0.6930 corrupt_frac=0.5402 loss_all=1.0561 loss_corrupt=1.8968 acc_corrupt_t_0p0_0p2=0.2058 corrupt_frac_t_0p0_0p2=0.2778 acc_corrupt_t_0p4_0p6=0.6782 corrupt_frac_t_0p4_0p6=0.1471 acc_corrupt_t_0p6_0p8=0.8676 corrupt_frac_t_0p6_0p8=0.1442 acc_corrupt_t_0p8_1p0=0.9539 corrupt_frac_t_0p8_1p0=0.4309 wrong_frac=0.3904 init_acc_corrupt=0.5841 init_gold_top10=0.6021 init_gold_top100=0.6101
+step=3200 micro_steps=25600 elapsed=127.6s lr=3.000000e-04 loss=2.6145 loss_recon=2.6145 loss_meanflow=0.0000 mean_model_t=0.4877 mean_corrupt_t=0.4877 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.7679 out_g_norm=0.0149 acc_all=0.7285 acc_corrupt=0.5476 corrupt_frac=0.5986 loss_all=1.7034 loss_corrupt=2.8148 acc_corrupt_t_0p0_0p2=0.1863 corrupt_frac_t_0p0_0p2=0.1882 acc_corrupt_t_0p2_0p4=0.4024 corrupt_frac_t_0p2_0p4=0.3281 acc_corrupt_t_0p4_0p6=0.6057 corrupt_frac_t_0p4_0p6=0.0608 acc_corrupt_t_0p6_0p8=0.7999 corrupt_frac_t_0p6_0p8=0.3608 acc_corrupt_t_0p8_1p0=0.8867 corrupt_frac_t_0p8_1p0=0.0621 wrong_frac=0.5645 init_acc_corrupt=0.3863 init_gold_top10=0.4295 init_gold_top100=0.4357
+step=3250 micro_steps=26000 elapsed=175.3s lr=3.000000e-04 loss=2.5833 loss_recon=2.5833 loss_meanflow=0.0000 mean_model_t=0.4957 mean_corrupt_t=0.4957 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.8014 out_g_norm=0.0148 acc_all=0.7147 acc_corrupt=0.5662 corrupt_frac=0.6557 loss_all=1.8117 loss_corrupt=2.7388 acc_corrupt_t_0p0_0p2=0.1201 corrupt_frac_t_0p0_0p2=0.1147 acc_corrupt_t_0p2_0p4=0.3546 corrupt_frac_t_0p2_0p4=0.3685 acc_corrupt_t_0p4_0p6=0.6875 corrupt_frac_t_0p4_0p6=0.1323 acc_corrupt_t_0p6_0p8=0.7612 corrupt_frac_t_0p6_0p8=0.1719 acc_corrupt_t_0p8_1p0=0.9405 corrupt_frac_t_0p8_1p0=0.2126 wrong_frac=0.5173 init_acc_corrupt=0.4363 init_gold_top10=0.4753 init_gold_top100=0.4837
+step=3300 micro_steps=26400 elapsed=127.6s lr=3.000000e-04 loss=2.4928 loss_recon=2.4928 loss_meanflow=0.0000 mean_model_t=0.5074 mean_corrupt_t=0.5074 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.8278 out_g_norm=0.0144 acc_all=0.6918 acc_corrupt=0.5090 corrupt_frac=0.6082 loss_all=1.8383 loss_corrupt=2.8961 acc_corrupt_t_0p0_0p2=0.1773 corrupt_frac_t_0p0_0p2=0.4835 acc_corrupt_t_0p2_0p4=0.5258 corrupt_frac_t_0p2_0p4=0.0584 acc_corrupt_t_0p4_0p6=0.6334 corrupt_frac_t_0p4_0p6=0.0734 acc_corrupt_t_0p6_0p8=0.7753 corrupt_frac_t_0p6_0p8=0.1537 acc_corrupt_t_0p8_1p0=0.9826 corrupt_frac_t_0p8_1p0=0.2310 wrong_frac=0.5950 init_acc_corrupt=0.3712 init_gold_top10=0.3926 init_gold_top100=0.4018
+step=3350 micro_steps=26800 elapsed=127.7s lr=3.000000e-04 loss=2.4775 loss_recon=2.4775 loss_meanflow=0.0000 mean_model_t=0.5006 mean_corrupt_t=0.5006 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.8604 out_g_norm=0.0155 acc_all=0.8364 acc_corrupt=0.6639 corrupt_frac=0.4865 loss_all=1.0054 loss_corrupt=2.0573 acc_corrupt_t_0p2_0p4=0.3736 corrupt_frac_t_0p2_0p4=0.1747 acc_corrupt_t_0p4_0p6=0.6234 corrupt_frac_t_0p4_0p6=0.5304 acc_corrupt_t_0p6_0p8=0.8596 corrupt_frac_t_0p6_0p8=0.1483 acc_corrupt_t_0p8_1p0=0.9581 corrupt_frac_t_0p8_1p0=0.1467 wrong_frac=0.4442 init_acc_corrupt=0.5365 init_gold_top10=0.5548 init_gold_top100=0.5575
+step=3400 micro_steps=27200 elapsed=127.8s lr=3.000000e-04 loss=2.5129 loss_recon=2.5129 loss_meanflow=0.0000 mean_model_t=0.4976 mean_corrupt_t=0.4976 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.8924 out_g_norm=0.0153 acc_all=0.8296 acc_corrupt=0.6626 corrupt_frac=0.4969 loss_all=1.0641 loss_corrupt=2.0784 acc_corrupt_t_0p0_0p2=0.3536 corrupt_frac_t_0p0_0p2=0.0709 acc_corrupt_t_0p2_0p4=0.3471 corrupt_frac_t_0p2_0p4=0.2056 acc_corrupt_t_0p4_0p6=0.6397 corrupt_frac_t_0p4_0p6=0.1670 acc_corrupt_t_0p6_0p8=0.7876 corrupt_frac_t_0p6_0p8=0.4036 acc_corrupt_t_0p8_1p0=0.9253 corrupt_frac_t_0p8_1p0=0.1529 wrong_frac=0.4425 init_acc_corrupt=0.5321 init_gold_top10=0.5549 init_gold_top100=0.5588
+step=3450 micro_steps=27600 elapsed=127.3s lr=3.000000e-04 loss=2.6037 loss_recon=2.6037 loss_meanflow=0.0000 mean_model_t=0.4975 mean_corrupt_t=0.4975 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.9291 out_g_norm=0.0141 acc_all=0.7703 acc_corrupt=0.6035 corrupt_frac=0.5694 loss_all=1.4758 loss_corrupt=2.5194 acc_corrupt_t_0p0_0p2=0.1498 corrupt_frac_t_0p0_0p2=0.2934 acc_corrupt_t_0p2_0p4=0.4774 corrupt_frac_t_0p2_0p4=0.0593 acc_corrupt_t_0p4_0p6=0.6851 corrupt_frac_t_0p4_0p6=0.1202 acc_corrupt_t_0p6_0p8=0.8229 corrupt_frac_t_0p6_0p8=0.3225 acc_corrupt_t_0p8_1p0=0.8968 corrupt_frac_t_0p8_1p0=0.2046 wrong_frac=0.4948 init_acc_corrupt=0.4857 init_gold_top10=0.4977 init_gold_top100=0.5043
+step=3500 micro_steps=28000 elapsed=127.8s lr=3.000000e-04 loss=2.4503 loss_recon=2.4503 loss_meanflow=0.0000 mean_model_t=0.5024 mean_corrupt_t=0.5024 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1437.9701 out_g_norm=0.0170 acc_all=0.6820 acc_corrupt=0.4972 corrupt_frac=0.6281 loss_all=1.9257 loss_corrupt=3.0231 acc_corrupt_t_0p0_0p2=0.2649 corrupt_frac_t_0p0_0p2=0.3137 acc_corrupt_t_0p2_0p4=0.4566 corrupt_frac_t_0p2_0p4=0.1656 acc_corrupt_t_0p4_0p6=0.6000 corrupt_frac_t_0p4_0p6=0.4062 acc_corrupt_t_0p6_0p8=0.7780 corrupt_frac_t_0p6_0p8=0.0884 acc_corrupt_t_0p8_1p0=0.9963 corrupt_frac_t_0p8_1p0=0.0260 wrong_frac=0.6489 init_acc_corrupt=0.3135 init_gold_top10=0.3424 init_gold_top100=0.3500
+step=3550 micro_steps=28400 elapsed=129.1s lr=3.000000e-04 loss=2.3950 loss_recon=2.3950 loss_meanflow=0.0000 mean_model_t=0.5046 mean_corrupt_t=0.5046 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.0044 out_g_norm=0.0150 acc_all=0.7578 acc_corrupt=0.6122 corrupt_frac=0.6200 loss_all=1.5253 loss_corrupt=2.4125 acc_corrupt_t_0p0_0p2=0.2389 corrupt_frac_t_0p0_0p2=0.2456 acc_corrupt_t_0p2_0p4=0.4583 corrupt_frac_t_0p2_0p4=0.0921 acc_corrupt_t_0p4_0p6=0.6300 corrupt_frac_t_0p4_0p6=0.2469 acc_corrupt_t_0p6_0p8=0.8142 corrupt_frac_t_0p6_0p8=0.2740 acc_corrupt_t_0p8_1p0=0.9387 corrupt_frac_t_0p8_1p0=0.1414 wrong_frac=0.4884 init_acc_corrupt=0.4836 init_gold_top10=0.5053 init_gold_top100=0.5121
+step=3600 micro_steps=28800 elapsed=191.8s lr=3.000000e-04 loss=2.4143 loss_recon=2.4143 loss_meanflow=0.0000 mean_model_t=0.5186 mean_corrupt_t=0.5186 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.0403 out_g_norm=0.0154 acc_all=0.6730 acc_corrupt=0.4876 corrupt_frac=0.6334 loss_all=2.0922 loss_corrupt=3.2643 acc_corrupt_t_0p0_0p2=0.1370 corrupt_frac_t_0p0_0p2=0.1934 acc_corrupt_t_0p2_0p4=0.3826 corrupt_frac_t_0p2_0p4=0.2418 acc_corrupt_t_0p4_0p6=0.6180 corrupt_frac_t_0p4_0p6=0.4735 acc_corrupt_t_0p6_0p8=0.8321 corrupt_frac_t_0p6_0p8=0.0913 wrong_frac=0.6238 init_acc_corrupt=0.3367 init_gold_top10=0.3711 init_gold_top100=0.3758
+step=3650 micro_steps=29200 elapsed=219.2s lr=3.000000e-04 loss=2.5185 loss_recon=2.5185 loss_meanflow=0.0000 mean_model_t=0.5013 mean_corrupt_t=0.5013 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.0772 out_g_norm=0.0149 acc_all=0.7538 acc_corrupt=0.5342 corrupt_frac=0.5274 loss_all=1.5719 loss_corrupt=2.9423 acc_corrupt_t_0p0_0p2=0.2060 corrupt_frac_t_0p0_0p2=0.3326 acc_corrupt_t_0p2_0p4=0.4605 corrupt_frac_t_0p2_0p4=0.1010 acc_corrupt_t_0p4_0p6=0.5873 corrupt_frac_t_0p4_0p6=0.2871 acc_corrupt_t_0p6_0p8=0.7724 corrupt_frac_t_0p6_0p8=0.0941 acc_corrupt_t_0p8_1p0=0.9606 corrupt_frac_t_0p8_1p0=0.1852 wrong_frac=0.5544 init_acc_corrupt=0.4079 init_gold_top10=0.4362 init_gold_top100=0.4462
+step=3700 micro_steps=29600 elapsed=179.9s lr=3.000000e-04 loss=2.4679 loss_recon=2.4679 loss_meanflow=0.0000 mean_model_t=0.5130 mean_corrupt_t=0.5130 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.1130 out_g_norm=0.0145 acc_all=0.8342 acc_corrupt=0.6832 corrupt_frac=0.5200 loss_all=1.0293 loss_corrupt=1.9436 acc_corrupt_t_0p0_0p2=0.1624 corrupt_frac_t_0p0_0p2=0.1684 acc_corrupt_t_0p4_0p6=0.6700 corrupt_frac_t_0p4_0p6=0.3038 acc_corrupt_t_0p6_0p8=0.7804 corrupt_frac_t_0p6_0p8=0.2646 acc_corrupt_t_0p8_1p0=0.9340 corrupt_frac_t_0p8_1p0=0.2632 wrong_frac=0.4094 init_acc_corrupt=0.5766 init_gold_top10=0.5862 init_gold_top100=0.5919
+step=3750 micro_steps=30000 elapsed=128.6s lr=3.000000e-04 loss=2.4525 loss_recon=2.4525 loss_meanflow=0.0000 mean_model_t=0.5041 mean_corrupt_t=0.5041 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.1452 out_g_norm=0.0137 acc_all=0.7866 acc_corrupt=0.5120 corrupt_frac=0.4318 loss_all=1.3797 loss_corrupt=3.1090 acc_corrupt_t_0p0_0p2=0.2392 corrupt_frac_t_0p0_0p2=0.2364 acc_corrupt_t_0p2_0p4=0.3762 corrupt_frac_t_0p2_0p4=0.3773 acc_corrupt_t_0p4_0p6=0.6735 corrupt_frac_t_0p4_0p6=0.0900 acc_corrupt_t_0p6_0p8=0.8067 corrupt_frac_t_0p6_0p8=0.1814 acc_corrupt_t_0p8_1p0=0.9274 corrupt_frac_t_0p8_1p0=0.1149 wrong_frac=0.6179 init_acc_corrupt=0.3264 init_gold_top10=0.3739 init_gold_top100=0.3822
+step=3800 micro_steps=30400 elapsed=128.3s lr=3.000000e-04 loss=2.4515 loss_recon=2.4515 loss_meanflow=0.0000 mean_model_t=0.5096 mean_corrupt_t=0.5096 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.1779 out_g_norm=0.0138 acc_all=0.8350 acc_corrupt=0.6069 corrupt_frac=0.4124 loss_all=0.9658 loss_corrupt=2.2753 acc_corrupt_t_0p0_0p2=0.4688 corrupt_frac_t_0p0_0p2=0.0142 acc_corrupt_t_0p2_0p4=0.4925 corrupt_frac_t_0p2_0p4=0.5935 acc_corrupt_t_0p4_0p6=0.6746 corrupt_frac_t_0p4_0p6=0.1178 acc_corrupt_t_0p6_0p8=0.8102 corrupt_frac_t_0p6_0p8=0.2402 acc_corrupt_t_0p8_1p0=0.9870 corrupt_frac_t_0p8_1p0=0.0342 wrong_frac=0.5426 init_acc_corrupt=0.4093 init_gold_top10=0.4550 init_gold_top100=0.4596
+step=3850 micro_steps=30800 elapsed=137.0s lr=3.000000e-04 loss=2.5001 loss_recon=2.5001 loss_meanflow=0.0000 mean_model_t=0.4848 mean_corrupt_t=0.4848 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.2111 out_g_norm=0.0139 acc_all=0.8232 acc_corrupt=0.6890 corrupt_frac=0.5645 loss_all=1.0883 loss_corrupt=1.8999 acc_corrupt_t_0p0_0p2=0.2523 corrupt_frac_t_0p0_0p2=0.0591 acc_corrupt_t_0p2_0p4=0.3512 corrupt_frac_t_0p2_0p4=0.1776 acc_corrupt_t_0p4_0p6=0.6347 corrupt_frac_t_0p4_0p6=0.2516 acc_corrupt_t_0p6_0p8=0.8263 corrupt_frac_t_0p6_0p8=0.2801 acc_corrupt_t_0p8_1p0=0.9528 corrupt_frac_t_0p8_1p0=0.2315 wrong_frac=0.4125 init_acc_corrupt=0.5681 init_gold_top10=0.5856 init_gold_top100=0.5876
+step=3900 micro_steps=31200 elapsed=275.9s lr=3.000000e-04 loss=2.5571 loss_recon=2.5571 loss_meanflow=0.0000 mean_model_t=0.4961 mean_corrupt_t=0.4961 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 out_w_norm=1438.2457 out_g_norm=0.0176 acc_all=0.8155 acc_corrupt=0.5732 corrupt_frac=0.4196 loss_all=1.1297 loss_corrupt=2.5263 acc_corrupt_t_0p0_0p2=0.2637 corrupt_frac_t_0p0_0p2=0.4127 acc_corrupt_t_0p2_0p4=0.5651 corrupt_frac_t_0p2_0p4=0.1341 acc_corrupt_t_0p4_0p6=0.6612 corrupt_frac_t_0p4_0p6=0.0447 acc_corrupt_t_0p6_0p8=0.8271 corrupt_frac_t_0p6_0p8=0.2600 acc_corrupt_t_0p8_1p0=0.9696 corrupt_frac_t_0p8_1p0=0.1485 wrong_frac=0.5870 init_acc_corrupt=0.3843 init_gold_top10=0.4018 init_gold_top100=0.4098
+W0514 15:05:53.653000 1231273 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
+W0514 15:05:53.654000 1231273 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 1231277 closing signal SIGTERM
+W0514 15:05:53.659000 1231273 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 1231278 closing signal SIGTERM
+W0514 15:05:53.660000 1231273 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 1231279 closing signal SIGTERM
+W0514 15:05:53.661000 1231273 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 1231280 closing signal SIGTERM
+Traceback (most recent call last):
+  File "<frozen runpy>", line 198, in _run_module_as_main
+  File "<frozen runpy>", line 88, in _run_code
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 922, in <module>
+    main()
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
+    return f(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 918, in main
+    run(args)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 909, in run
+    elastic_launch(
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
+    return launch_agent(self._config, self._entrypoint, list(args))
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
+    result = agent.run()
+             ^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 137, in wrapper
+    result = f(*args, **kwargs)
+             ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
+    result = self._invoke_run(role)
+             ^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
+    time.sleep(monitor_interval)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
+    raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
+torch.distributed.elastic.multiprocessing.api.SignalException: Process 1231273 got signal: 15

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/tests/test_numpy_version.py

@@ -0,0 +1,41 @@
+"""
+Check the numpy version is valid.
+
+Note that a development version is marked by the presence of 'dev0' or '+'
+in the version string, all else is treated as a release. The version string
+itself is set from the output of ``git describe`` which relies on tags.
+
+Examples
+--------
+
+Valid Development: 1.22.0.dev0 1.22.0.dev0+5-g7999db4df2 1.22.0+5-g7999db4df2
+Valid Release: 1.21.0.rc1, 1.21.0.b1, 1.21.0
+Invalid: 1.22.0.dev, 1.22.0.dev0-5-g7999db4dfB, 1.21.0.d1, 1.21.a
+
+Note that a release is determined by the version string, which in turn
+is controlled by the result of the ``git describe`` command.
+"""
+import re
+
+import numpy as np
+from numpy.testing import assert_
+
+
+def test_valid_numpy_version():
+    # Verify that the numpy version is a valid one (no .post suffix or other
+    # nonsense).  See gh-6431 for an issue caused by an invalid version.
+    version_pattern = r"^[0-9]+\.[0-9]+\.[0-9]+(a[0-9]|b[0-9]|rc[0-9])?"
+    dev_suffix = r"(\.dev[0-9]+(\+git[0-9]+\.[0-9a-f]+)?)?"
+    res = re.match(version_pattern + dev_suffix + '$', np.__version__)
+
+    assert_(res is not None, np.__version__)
+
+
+def test_short_version():
+    # Check numpy.short_version actually exists
+    if np.version.release:
+        assert_(np.__version__ == np.version.short_version,
+                "short_version mismatch in release version")
+    else:
+        assert_(np.__version__.split("+")[0] == np.version.short_version,
+                "short_version mismatch in development version")

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/tests/test_public_api.py

@@ -0,0 +1,551 @@
+import sys
+import sysconfig
+import subprocess
+import pkgutil
+import types
+import importlib
+import warnings
+
+import numpy as np
+import numpy
+import pytest
+from numpy.testing import IS_WASM
+
+try:
+    import ctypes
+except ImportError:
+    ctypes = None
+
+
+def check_dir(module, module_name=None):
+    """Returns a mapping of all objects with the wrong __module__ attribute."""
+    if module_name is None:
+        module_name = module.__name__
+    results = {}
+    for name in dir(module):
+        item = getattr(module, name)
+        if (hasattr(item, '__module__') and hasattr(item, '__name__')
+                and item.__module__ != module_name):
+            results[name] = item.__module__ + '.' + item.__name__
+    return results
+
+
+def test_numpy_namespace():
+    # None of these objects are publicly documented to be part of the main
+    # NumPy namespace (some are useful though, others need to be cleaned up)
+    undocumented = {
+        '_add_newdoc_ufunc': 'numpy.core._multiarray_umath._add_newdoc_ufunc',
+        'add_docstring': 'numpy.core._multiarray_umath.add_docstring',
+        'add_newdoc': 'numpy.core.function_base.add_newdoc',
+        'add_newdoc_ufunc': 'numpy.core._multiarray_umath._add_newdoc_ufunc',
+        'byte_bounds': 'numpy.lib.utils.byte_bounds',
+        'compare_chararrays': 'numpy.core._multiarray_umath.compare_chararrays',
+        'deprecate': 'numpy.lib.utils.deprecate',
+        'deprecate_with_doc': 'numpy.lib.utils.deprecate_with_doc',
+        'disp': 'numpy.lib.function_base.disp',
+        'fastCopyAndTranspose': 'numpy.core._multiarray_umath.fastCopyAndTranspose',
+        'get_array_wrap': 'numpy.lib.shape_base.get_array_wrap',
+        'get_include': 'numpy.lib.utils.get_include',
+        'recfromcsv': 'numpy.lib.npyio.recfromcsv',
+        'recfromtxt': 'numpy.lib.npyio.recfromtxt',
+        'safe_eval': 'numpy.lib.utils.safe_eval',
+        'set_string_function': 'numpy.core.arrayprint.set_string_function',
+        'show_config': 'numpy.__config__.show',
+        'show_runtime': 'numpy.lib.utils.show_runtime',
+        'who': 'numpy.lib.utils.who',
+    }
+    # We override dir to not show these members
+    allowlist = undocumented
+    bad_results = check_dir(np)
+    # pytest gives better error messages with the builtin assert than with
+    # assert_equal
+    assert bad_results == allowlist
+
+
+@pytest.mark.skipif(IS_WASM, reason="can't start subprocess")
+@pytest.mark.parametrize('name', ['testing'])
+def test_import_lazy_import(name):
+    """Make sure we can actually use the modules we lazy load.
+
+    While not exported as part of the public API, it was accessible.  With the
+    use of __getattr__ and __dir__, this isn't always true It can happen that
+    an infinite recursion may happen.
+
+    This is the only way I found that would force the failure to appear on the
+    badly implemented code.
+
+    We also test for the presence of the lazily imported modules in dir
+
+    """
+    exe = (sys.executable, '-c', "import numpy; numpy." + name)
+    result = subprocess.check_output(exe)
+    assert not result
+
+    # Make sure they are still in the __dir__
+    assert name in dir(np)
+
+
+def test_dir_testing():
+    """Assert that output of dir has only one "testing/tester"
+    attribute without duplicate"""
+    assert len(dir(np)) == len(set(dir(np)))
+
+
+def test_numpy_linalg():
+    bad_results = check_dir(np.linalg)
+    assert bad_results == {}
+
+
+def test_numpy_fft():
+    bad_results = check_dir(np.fft)
+    assert bad_results == {}
+
+
+@pytest.mark.skipif(ctypes is None,
+                    reason="ctypes not available in this python")
+def test_NPY_NO_EXPORT():
+    cdll = ctypes.CDLL(np.core._multiarray_tests.__file__)
+    # Make sure an arbitrary NPY_NO_EXPORT function is actually hidden
+    f = getattr(cdll, 'test_not_exported', None)
+    assert f is None, ("'test_not_exported' is mistakenly exported, "
+                      "NPY_NO_EXPORT does not work")
+
+
+# Historically NumPy has not used leading underscores for private submodules
+# much.  This has resulted in lots of things that look like public modules
+# (i.e. things that can be imported as `import numpy.somesubmodule.somefile`),
+# but were never intended to be public.  The PUBLIC_MODULES list contains
+# modules that are either public because they were meant to be, or because they
+# contain public functions/objects that aren't present in any other namespace
+# for whatever reason and therefore should be treated as public.
+#
+# The PRIVATE_BUT_PRESENT_MODULES list contains modules that look public (lack
+# of underscores) but should not be used.  For many of those modules the
+# current status is fine.  For others it may make sense to work on making them
+# private, to clean up our public API and avoid confusion.
+PUBLIC_MODULES = ['numpy.' + s for s in [
+    "array_api",
+    "array_api.linalg",
+    "ctypeslib",
+    "doc",
+    "doc.constants",
+    "doc.ufuncs",
+    "dtypes",
+    "exceptions",
+    "f2py",
+    "fft",
+    "lib",
+    "lib.format",  # was this meant to be public?
+    "lib.mixins",
+    "lib.recfunctions",
+    "lib.scimath",
+    "lib.stride_tricks",
+    "linalg",
+    "ma",
+    "ma.extras",
+    "ma.mrecords",
+    "matlib",
+    "polynomial",
+    "polynomial.chebyshev",
+    "polynomial.hermite",
+    "polynomial.hermite_e",
+    "polynomial.laguerre",
+    "polynomial.legendre",
+    "polynomial.polynomial",
+    "random",
+    "testing",
+    "testing.overrides",
+    "typing",
+    "typing.mypy_plugin",
+    "version"  # Should be removed for NumPy 2.0
+]]
+if sys.version_info < (3, 12):
+    PUBLIC_MODULES += [
+        'numpy.' + s for s in [
+            "distutils",
+            "distutils.cpuinfo",
+            "distutils.exec_command",
+            "distutils.misc_util",
+            "distutils.log",
+            "distutils.system_info",
+        ]
+    ]
+
+
+
+PUBLIC_ALIASED_MODULES = [
+    "numpy.char",
+    "numpy.emath",
+    "numpy.rec",
+]
+
+
+PRIVATE_BUT_PRESENT_MODULES = ['numpy.' + s for s in [
+    "compat",
+    "compat.py3k",
+    "conftest",
+    "core",
+    "core.arrayprint",
+    "core.defchararray",
+    "core.einsumfunc",
+    "core.fromnumeric",
+    "core.function_base",
+    "core.getlimits",
+    "core.memmap",
+    "core.multiarray",
+    "core.numeric",
+    "core.numerictypes",
+    "core.overrides",
+    "core.records",
+    "core.shape_base",
+    "core.umath",
+    "f2py.auxfuncs",
+    "f2py.capi_maps",
+    "f2py.cb_rules",
+    "f2py.cfuncs",
+    "f2py.common_rules",
+    "f2py.crackfortran",
+    "f2py.diagnose",
+    "f2py.f2py2e",
+    "f2py.f90mod_rules",
+    "f2py.func2subr",
+    "f2py.rules",
+    "f2py.symbolic",
+    "f2py.use_rules",
+    "fft.helper",
+    "lib.arraypad",
+    "lib.arraysetops",
+    "lib.arrayterator",
+    "lib.function_base",
+    "lib.histograms",
+    "lib.index_tricks",
+    "lib.nanfunctions",
+    "lib.npyio",
+    "lib.polynomial",
+    "lib.shape_base",
+    "lib.twodim_base",
+    "lib.type_check",
+    "lib.ufunclike",
+    "lib.user_array",  # note: not in np.lib, but probably should just be deleted
+    "lib.utils",
+    "linalg.lapack_lite",
+    "linalg.linalg",
+    "ma.core",
+    "ma.testutils",
+    "ma.timer_comparison",
+    "matrixlib",
+    "matrixlib.defmatrix",
+    "polynomial.polyutils",
+    "random.mtrand",
+    "random.bit_generator",
+    "testing.print_coercion_tables",
+]]
+if sys.version_info < (3, 12):
+    PRIVATE_BUT_PRESENT_MODULES += [
+        'numpy.' + s for s in [
+            "distutils.armccompiler",
+            "distutils.fujitsuccompiler",
+            "distutils.ccompiler",
+            'distutils.ccompiler_opt',
+            "distutils.command",
+            "distutils.command.autodist",
+            "distutils.command.bdist_rpm",
+            "distutils.command.build",
+            "distutils.command.build_clib",
+            "distutils.command.build_ext",
+            "distutils.command.build_py",
+            "distutils.command.build_scripts",
+            "distutils.command.build_src",
+            "distutils.command.config",
+            "distutils.command.config_compiler",
+            "distutils.command.develop",
+            "distutils.command.egg_info",
+            "distutils.command.install",
+            "distutils.command.install_clib",
+            "distutils.command.install_data",
+            "distutils.command.install_headers",
+            "distutils.command.sdist",
+            "distutils.conv_template",
+            "distutils.core",
+            "distutils.extension",
+            "distutils.fcompiler",
+            "distutils.fcompiler.absoft",
+            "distutils.fcompiler.arm",
+            "distutils.fcompiler.compaq",
+            "distutils.fcompiler.environment",
+            "distutils.fcompiler.g95",
+            "distutils.fcompiler.gnu",
+            "distutils.fcompiler.hpux",
+            "distutils.fcompiler.ibm",
+            "distutils.fcompiler.intel",
+            "distutils.fcompiler.lahey",
+            "distutils.fcompiler.mips",
+            "distutils.fcompiler.nag",
+            "distutils.fcompiler.none",
+            "distutils.fcompiler.pathf95",
+            "distutils.fcompiler.pg",
+            "distutils.fcompiler.nv",
+            "distutils.fcompiler.sun",
+            "distutils.fcompiler.vast",
+            "distutils.fcompiler.fujitsu",
+            "distutils.from_template",
+            "distutils.intelccompiler",
+            "distutils.lib2def",
+            "distutils.line_endings",
+            "distutils.mingw32ccompiler",
+            "distutils.msvccompiler",
+            "distutils.npy_pkg_config",
+            "distutils.numpy_distribution",
+            "distutils.pathccompiler",
+            "distutils.unixccompiler",
+        ]
+    ]
+
+
+def is_unexpected(name):
+    """Check if this needs to be considered."""
+    if '._' in name or '.tests' in name or '.setup' in name:
+        return False
+
+    if name in PUBLIC_MODULES:
+        return False
+
+    if name in PUBLIC_ALIASED_MODULES:
+        return False
+
+    if name in PRIVATE_BUT_PRESENT_MODULES:
+        return False
+
+    return True
+
+
+# These are present in a directory with an __init__.py but cannot be imported
+# code_generators/ isn't installed, but present for an inplace build
+SKIP_LIST = [
+    "numpy.core.code_generators",
+    "numpy.core.code_generators.genapi",
+    "numpy.core.code_generators.generate_umath",
+    "numpy.core.code_generators.ufunc_docstrings",
+    "numpy.core.code_generators.generate_numpy_api",
+    "numpy.core.code_generators.generate_ufunc_api",
+    "numpy.core.code_generators.numpy_api",
+    "numpy.core.code_generators.generate_umath_doc",
+    "numpy.core.code_generators.verify_c_api_version",
+    "numpy.core.cversions",
+    "numpy.core.generate_numpy_api",
+    "numpy.core.umath_tests",
+]
+if sys.version_info < (3, 12):
+    SKIP_LIST += ["numpy.distutils.msvc9compiler"]
+
+
+# suppressing warnings from deprecated modules
+@pytest.mark.filterwarnings("ignore:.*np.compat.*:DeprecationWarning")
+def test_all_modules_are_expected():
+    """
+    Test that we don't add anything that looks like a new public module by
+    accident.  Check is based on filenames.
+    """
+
+    modnames = []
+    for _, modname, ispkg in pkgutil.walk_packages(path=np.__path__,
+                                                   prefix=np.__name__ + '.',
+                                                   onerror=None):
+        if is_unexpected(modname) and modname not in SKIP_LIST:
+            # We have a name that is new.  If that's on purpose, add it to
+            # PUBLIC_MODULES.  We don't expect to have to add anything to
+            # PRIVATE_BUT_PRESENT_MODULES.  Use an underscore in the name!
+            modnames.append(modname)
+
+    if modnames:
+        raise AssertionError(f'Found unexpected modules: {modnames}')
+
+
+# Stuff that clearly shouldn't be in the API and is detected by the next test
+# below
+SKIP_LIST_2 = [
+    'numpy.math',
+    'numpy.doc.constants.re',
+    'numpy.doc.constants.textwrap',
+    'numpy.lib.emath',
+    'numpy.lib.math',
+    'numpy.matlib.char',
+    'numpy.matlib.rec',
+    'numpy.matlib.emath',
+    'numpy.matlib.exceptions',
+    'numpy.matlib.math',
+    'numpy.matlib.linalg',
+    'numpy.matlib.fft',
+    'numpy.matlib.random',
+    'numpy.matlib.ctypeslib',
+    'numpy.matlib.ma',
+]
+if sys.version_info < (3, 12):
+    SKIP_LIST_2 += [
+        'numpy.distutils.log.sys',
+        'numpy.distutils.log.logging',
+        'numpy.distutils.log.warnings',
+    ]
+
+
+def test_all_modules_are_expected_2():
+    """
+    Method checking all objects. The pkgutil-based method in
+    `test_all_modules_are_expected` does not catch imports into a namespace,
+    only filenames.  So this test is more thorough, and checks this like:
+
+        import .lib.scimath as emath
+
+    To check if something in a module is (effectively) public, one can check if
+    there's anything in that namespace that's a public function/object but is
+    not exposed in a higher-level namespace.  For example for a `numpy.lib`
+    submodule::
+
+        mod = np.lib.mixins
+        for obj in mod.__all__:
+            if obj in np.__all__:
+                continue
+            elif obj in np.lib.__all__:
+                continue
+
+            else:
+                print(obj)
+
+    """
+
+    def find_unexpected_members(mod_name):
+        members = []
+        module = importlib.import_module(mod_name)
+        if hasattr(module, '__all__'):
+            objnames = module.__all__
+        else:
+            objnames = dir(module)
+
+        for objname in objnames:
+            if not objname.startswith('_'):
+                fullobjname = mod_name + '.' + objname
+                if isinstance(getattr(module, objname), types.ModuleType):
+                    if is_unexpected(fullobjname):
+                        if fullobjname not in SKIP_LIST_2:
+                            members.append(fullobjname)
+
+        return members
+
+    unexpected_members = find_unexpected_members("numpy")
+    for modname in PUBLIC_MODULES:
+        unexpected_members.extend(find_unexpected_members(modname))
+
+    if unexpected_members:
+        raise AssertionError("Found unexpected object(s) that look like "
+                             "modules: {}".format(unexpected_members))
+
+
+def test_api_importable():
+    """
+    Check that all submodules listed higher up in this file can be imported
+
+    Note that if a PRIVATE_BUT_PRESENT_MODULES entry goes missing, it may
+    simply need to be removed from the list (deprecation may or may not be
+    needed - apply common sense).
+    """
+    def check_importable(module_name):
+        try:
+            importlib.import_module(module_name)
+        except (ImportError, AttributeError):
+            return False
+
+        return True
+
+    module_names = []
+    for module_name in PUBLIC_MODULES:
+        if not check_importable(module_name):
+            module_names.append(module_name)
+
+    if module_names:
+        raise AssertionError("Modules in the public API that cannot be "
+                             "imported: {}".format(module_names))
+
+    for module_name in PUBLIC_ALIASED_MODULES:
+        try:
+            eval(module_name)
+        except AttributeError:
+            module_names.append(module_name)
+
+    if module_names:
+        raise AssertionError("Modules in the public API that were not "
+                             "found: {}".format(module_names))
+
+    with warnings.catch_warnings(record=True) as w:
+        warnings.filterwarnings('always', category=DeprecationWarning)
+        warnings.filterwarnings('always', category=ImportWarning)
+        for module_name in PRIVATE_BUT_PRESENT_MODULES:
+            if not check_importable(module_name):
+                module_names.append(module_name)
+
+    if module_names:
+        raise AssertionError("Modules that are not really public but looked "
+                             "public and can not be imported: "
+                             "{}".format(module_names))
+
+
+@pytest.mark.xfail(
+    sysconfig.get_config_var("Py_DEBUG") not in (None, 0, "0"),
+    reason=(
+        "NumPy possibly built with `USE_DEBUG=True ./tools/travis-test.sh`, "
+        "which does not expose the `array_api` entry point. "
+        "See https://github.com/numpy/numpy/pull/19800"
+    ),
+)
+def test_array_api_entry_point():
+    """
+    Entry point for Array API implementation can be found with importlib and
+    returns the numpy.array_api namespace.
+    """
+    # For a development install that did not go through meson-python,
+    # the entrypoint will not have been installed. So ensure this test fails
+    # only if numpy is inside site-packages.
+    numpy_in_sitepackages = sysconfig.get_path('platlib') in np.__file__
+
+    eps = importlib.metadata.entry_points()
+    try:
+        xp_eps = eps.select(group="array_api")
+    except AttributeError:
+        # The select interface for entry_points was introduced in py3.10,
+        # deprecating its dict interface. We fallback to dict keys for finding
+        # Array API entry points so that running this test in <=3.9 will
+        # still work - see https://github.com/numpy/numpy/pull/19800.
+        xp_eps = eps.get("array_api", [])
+    if len(xp_eps) == 0:
+        if numpy_in_sitepackages:
+            msg = "No entry points for 'array_api' found"
+            raise AssertionError(msg) from None
+        return
+
+    try:
+        ep = next(ep for ep in xp_eps if ep.name == "numpy")
+    except StopIteration:
+        if numpy_in_sitepackages:
+            msg = "'numpy' not in array_api entry points"
+            raise AssertionError(msg) from None
+        return
+
+    xp = ep.load()
+    msg = (
+        f"numpy entry point value '{ep.value}' "
+        "does not point to our Array API implementation"
+    )
+    assert xp is numpy.array_api, msg
+
+
+@pytest.mark.parametrize("name", [
+        'ModuleDeprecationWarning', 'VisibleDeprecationWarning',
+        'ComplexWarning', 'TooHardError', 'AxisError'])
+def test_moved_exceptions(name):
+    # These were moved to the exceptions namespace, but currently still
+    # available
+    assert name in np.__all__
+    assert name not in np.__dir__()
+    # Fetching works, but __module__ is set correctly:
+    assert getattr(np, name).__module__ == "numpy.exceptions"
+    assert name in np.exceptions.__all__
+    getattr(np.exceptions, name)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/tests/test_scripts.py

@@ -0,0 +1,47 @@
+""" Test scripts
+
+Test that we can run executable scripts that have been installed with numpy.
+"""
+import sys
+import os
+import pytest
+from os.path import join as pathjoin, isfile, dirname
+import subprocess
+
+import numpy as np
+from numpy.testing import assert_equal, IS_WASM
+
+is_inplace = isfile(pathjoin(dirname(np.__file__),  '..', 'setup.py'))
+
+
+def find_f2py_commands():
+    if sys.platform == 'win32':
+        exe_dir = dirname(sys.executable)
+        if exe_dir.endswith('Scripts'): # virtualenv
+            return [os.path.join(exe_dir, 'f2py')]
+        else:
+            return [os.path.join(exe_dir, "Scripts", 'f2py')]
+    else:
+        # Three scripts are installed in Unix-like systems:
+        # 'f2py', 'f2py{major}', and 'f2py{major.minor}'. For example,
+        # if installed with python3.9 the scripts would be named
+        # 'f2py', 'f2py3', and 'f2py3.9'.
+        version = sys.version_info
+        major = str(version.major)
+        minor = str(version.minor)
+        return ['f2py', 'f2py' + major, 'f2py' + major + '.' + minor]
+
+
+@pytest.mark.skipif(is_inplace, reason="Cannot test f2py command inplace")
+@pytest.mark.xfail(reason="Test is unreliable")
+@pytest.mark.parametrize('f2py_cmd', find_f2py_commands())
+def test_f2py(f2py_cmd):
+    # test that we can run f2py script
+    stdout = subprocess.check_output([f2py_cmd, '-v'])
+    assert_equal(stdout.strip(), np.__version__.encode('ascii'))
+
+
+@pytest.mark.skipif(IS_WASM, reason="Cannot start subprocess")
+def test_pep338():
+    stdout = subprocess.check_output([sys.executable, '-mnumpy.f2py', '-v'])
+    assert_equal(stdout.strip(), np.__version__.encode('ascii'))

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/superpoint/__init__.py

@@ -0,0 +1,29 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import _LazyModule
+from ...utils.import_utils import define_import_structure
+
+
+if TYPE_CHECKING:
+    from .configuration_superpoint import *
+    from .image_processing_pil_superpoint import *
+    from .image_processing_superpoint import *
+    from .modeling_superpoint import *
+else:
+    import sys
+
+    _file = globals()["__file__"]
+    sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/superpoint/configuration_superpoint.py

@@ -0,0 +1,66 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from huggingface_hub.dataclasses import strict
+
+from ...configuration_utils import PreTrainedConfig
+from ...utils import auto_docstring
+
+
+@auto_docstring(checkpoint="magic-leap-community/superpoint")
+@strict
+class SuperPointConfig(PreTrainedConfig):
+    r"""
+    encoder_hidden_sizes (`List`, *optional*, defaults to `[64, 64, 128, 128]`):
+        The number of channels in each convolutional layer in the encoder.
+    keypoint_decoder_dim (`int`, *optional*, defaults to 65):
+        The output dimension of the keypoint decoder.
+    descriptor_decoder_dim (`int`, *optional*, defaults to 256):
+        The output dimension of the descriptor decoder.
+    keypoint_threshold (`float`, *optional*, defaults to 0.005):
+        The threshold to use for extracting keypoints.
+    max_keypoints (`int`, *optional*, defaults to -1):
+        The maximum number of keypoints to extract. If `-1`, will extract all keypoints.
+    nms_radius (`int`, *optional*, defaults to 4):
+        The radius for non-maximum suppression.
+    border_removal_distance (`int`, *optional*, defaults to 4):
+        The distance from the border to remove keypoints.
+
+    Example:
+    ```python
+    >>> from transformers import SuperPointConfig, SuperPointForKeypointDetection
+
+    >>> # Initializing a SuperPoint superpoint style configuration
+    >>> configuration = SuperPointConfig()
+    >>> # Initializing a model from the superpoint style configuration
+    >>> model = SuperPointForKeypointDetection(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "superpoint"
+
+    encoder_hidden_sizes: list[int] | tuple[int, ...] = (64, 64, 128, 128)
+    decoder_hidden_size: int = 256
+    keypoint_decoder_dim: int = 65
+    descriptor_decoder_dim: int = 256
+    keypoint_threshold: float = 0.005
+    max_keypoints: int = -1
+    nms_radius: int = 4
+    border_removal_distance: int = 4
+    initializer_range: float = 0.02
+
+
+__all__ = ["SuperPointConfig"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/superpoint/image_processing_superpoint.py

@@ -0,0 +1,164 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for SuperPoint."""
+
+from typing import TYPE_CHECKING
+
+import torch
+
+from ...image_processing_backends import TorchvisionBackend
+from ...image_processing_utils import BatchFeature
+from ...image_transforms import group_images_by_shape, reorder_images
+from ...image_utils import PILImageResampling, SizeDict
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring
+
+
+if TYPE_CHECKING:
+    from .modeling_superpoint import SuperPointKeypointDescriptionOutput
+
+from torchvision.transforms.v2 import functional as tvF
+
+
+class SuperPointImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    do_grayscale (`bool`, *optional*, defaults to `self.do_grayscale`):
+        Whether to convert the image to grayscale. Can be overridden by `do_grayscale` in the `preprocess` method.
+    """
+
+    do_grayscale: bool
+
+
+def is_grayscale(image: "torch.Tensor") -> bool:
+    """Checks if an image is grayscale (all RGB channels are identical)."""
+    if image.ndim < 3 or image.shape[0 if image.ndim == 3 else 1] == 1:
+        return True
+    return torch.all(image[..., 0, :, :] == image[..., 1, :, :]) and torch.all(
+        image[..., 1, :, :] == image[..., 2, :, :]
+    )
+
+
+def convert_to_grayscale(image: "torch.Tensor") -> "torch.Tensor":
+    """
+    Converts an image to grayscale format using the NTSC formula. Only support torch.Tensor.
+
+    This function is supposed to return a 1-channel image, but it returns a 3-channel image with the same value in each
+    channel, because of an issue that is discussed in :
+    https://github.com/huggingface/transformers/pull/25786#issuecomment-1730176446
+
+    Args:
+        image (torch.Tensor):
+            The image to convert.
+    """
+    if is_grayscale(image):
+        return image
+    return tvF.rgb_to_grayscale(image, num_output_channels=3)
+
+
+@auto_docstring
+class SuperPointImageProcessor(TorchvisionBackend):
+    valid_kwargs = SuperPointImageProcessorKwargs
+    resample = PILImageResampling.BILINEAR
+    size = {"height": 480, "width": 640}
+    default_to_square = False
+    do_resize = True
+    do_rescale = True
+    rescale_factor = 1 / 255
+    do_normalize = None
+    do_grayscale = False
+
+    def __init__(self, **kwargs: Unpack[SuperPointImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        resample: "PILImageResampling | tvF.InterpolationMode | int | None",
+        do_rescale: bool,
+        rescale_factor: float,
+        disable_grouping: bool | None,
+        return_tensors: str | TensorType | None,
+        do_grayscale: bool = False,
+        **kwargs,
+    ) -> BatchFeature:
+        # Group images by size for batched processing
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            # Apply grayscale conversion before resize (if requested)
+            if do_grayscale:
+                stacked_images = convert_to_grayscale(stacked_images)
+            # Resize
+            if do_resize:
+                stacked_images = self.resize(stacked_images, size=size, resample=resample)
+            # Rescale
+            if do_rescale:
+                stacked_images = self.rescale(stacked_images, rescale_factor)
+            processed_images_grouped[shape] = stacked_images
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
+
+    def post_process_keypoint_detection(
+        self, outputs: "SuperPointKeypointDescriptionOutput", target_sizes: TensorType | list[tuple]
+    ) -> list[dict[str, "torch.Tensor"]]:
+        """
+        Converts the raw output of [`SuperPointForKeypointDetection`] into lists of keypoints, scores and descriptors
+        with coordinates absolute to the original image sizes.
+
+        Args:
+            outputs ([`SuperPointKeypointDescriptionOutput`]):
+                Raw outputs of the model containing keypoints in a relative (x, y) format, with scores and descriptors.
+            target_sizes (`torch.Tensor` or `list[tuple[int, int]]`):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`tuple[int, int]`) containing the target size
+                `(height, width)` of each image in the batch. This must be the original
+                image size (before any processing).
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the keypoints in absolute format according
+            to target_sizes, scores and descriptors for an image in the batch as predicted by the model.
+        """
+        if len(outputs.mask) != len(target_sizes):
+            raise ValueError("Make sure that you pass in as many target sizes as the batch dimension of the mask")
+
+        if isinstance(target_sizes, list):
+            image_sizes = torch.tensor(target_sizes, device=outputs.mask.device)
+        else:
+            if target_sizes.shape[1] != 2:
+                raise ValueError(
+                    "Each element of target_sizes must contain the size (h, w) of each image of the batch"
+                )
+            image_sizes = target_sizes
+
+        # Flip the image sizes to (width, height) and convert keypoints to absolute coordinates
+        image_sizes = torch.flip(image_sizes, [1])
+        masked_keypoints = outputs.keypoints * image_sizes[:, None]
+
+        # Convert masked_keypoints to int
+        masked_keypoints = masked_keypoints.to(torch.int32)
+
+        results = []
+        for image_mask, keypoints, scores, descriptors in zip(
+            outputs.mask, masked_keypoints, outputs.scores, outputs.descriptors
+        ):
+            indices = torch.nonzero(image_mask).squeeze(1)
+            keypoints = keypoints[indices]
+            scores = scores[indices]
+            descriptors = descriptors[indices]
+            results.append({"keypoints": keypoints, "scores": scores, "descriptors": descriptors})
+
+        return results
+
+
+__all__ = ["SuperPointImageProcessor"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/transformers/models/superpoint/modeling_superpoint.py

@@ -0,0 +1,468 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch SuperPoint model."""
+
+from dataclasses import dataclass
+
+import torch
+from torch import nn
+
+from transformers import PreTrainedModel
+from transformers.modeling_outputs import (
+    BaseModelOutputWithNoAttention,
+)
+from transformers.models.superpoint.configuration_superpoint import SuperPointConfig
+
+from ...utils import (
+    ModelOutput,
+    auto_docstring,
+    logging,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+def remove_keypoints_from_borders(
+    keypoints: torch.Tensor, scores: torch.Tensor, border: int, height: int, width: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Removes keypoints (and their associated scores) that are too close to the border"""
+    mask_h = (keypoints[:, 0] >= border) & (keypoints[:, 0] < (height - border))
+    mask_w = (keypoints[:, 1] >= border) & (keypoints[:, 1] < (width - border))
+    mask = mask_h & mask_w
+    return keypoints[mask], scores[mask]
+
+
+def top_k_keypoints(keypoints: torch.Tensor, scores: torch.Tensor, k: int) -> tuple[torch.Tensor, torch.Tensor]:
+    """Keeps the k keypoints with highest score"""
+    if k >= len(keypoints):
+        return keypoints, scores
+    scores, indices = torch.topk(scores, k, dim=0)
+    return keypoints[indices], scores
+
+
+def simple_nms(scores: torch.Tensor, nms_radius: int) -> torch.Tensor:
+    """Applies non-maximum suppression on scores"""
+    if nms_radius < 0:
+        raise ValueError("Expected positive values for nms_radius")
+
+    def max_pool(x):
+        return nn.functional.max_pool2d(x, kernel_size=nms_radius * 2 + 1, stride=1, padding=nms_radius)
+
+    zeros = torch.zeros_like(scores)
+    max_mask = scores == max_pool(scores)
+    for _ in range(2):
+        supp_mask = max_pool(max_mask.float()) > 0
+        supp_scores = torch.where(supp_mask, zeros, scores)
+        new_max_mask = supp_scores == max_pool(supp_scores)
+        max_mask = max_mask | (new_max_mask & (~supp_mask))
+    return torch.where(max_mask, scores, zeros)
+
+
+@auto_docstring(
+    custom_intro="""
+    Base class for outputs of image point description models. Due to the nature of keypoint detection, the number of
+    keypoints is not fixed and can vary from image to image, which makes batching non-trivial. In the batch of images,
+    the maximum number of keypoints is set as the dimension of the keypoints, scores and descriptors tensors. The mask
+    tensor is used to indicate which values in the keypoints, scores and descriptors tensors are keypoint information
+    and which are padding.
+    """
+)
+@dataclass
+class SuperPointKeypointDescriptionOutput(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*):
+        Loss computed during training.
+    keypoints (`torch.FloatTensor` of shape `(batch_size, num_keypoints, 2)`):
+        Relative (x, y) coordinates of predicted keypoints in a given image.
+    scores (`torch.FloatTensor` of shape `(batch_size, num_keypoints)`):
+        Scores of predicted keypoints.
+    descriptors (`torch.FloatTensor` of shape `(batch_size, num_keypoints, descriptor_size)`):
+        Descriptors of predicted keypoints.
+    mask (`torch.BoolTensor` of shape `(batch_size, num_keypoints)`):
+        Mask indicating which values in keypoints, scores and descriptors are keypoint information.
+    hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or
+    when `config.output_hidden_states=True`):
+        Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+        one for the output of each stage) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states
+        (also called feature maps) of the model at the output of each stage.
+    """
+
+    loss: torch.FloatTensor | None = None
+    keypoints: torch.IntTensor | None = None
+    scores: torch.FloatTensor | None = None
+    descriptors: torch.FloatTensor | None = None
+    mask: torch.BoolTensor | None = None
+    hidden_states: tuple[torch.FloatTensor] | None = None
+
+
+class SuperPointConvBlock(nn.Module):
+    def __init__(
+        self, config: SuperPointConfig, in_channels: int, out_channels: int, add_pooling: bool = False
+    ) -> None:
+        super().__init__()
+        self.conv_a = nn.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+        self.conv_b = nn.Conv2d(
+            out_channels,
+            out_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+        self.relu = nn.ReLU(inplace=True)
+        self.pool = nn.MaxPool2d(kernel_size=2, stride=2) if add_pooling else None
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.relu(self.conv_a(hidden_states))
+        hidden_states = self.relu(self.conv_b(hidden_states))
+        if self.pool is not None:
+            hidden_states = self.pool(hidden_states)
+        return hidden_states
+
+
+class SuperPointEncoder(nn.Module):
+    """
+    SuperPoint encoder module. It is made of 4 convolutional layers with ReLU activation and max pooling, reducing the
+     dimensionality of the image.
+    """
+
+    def __init__(self, config: SuperPointConfig) -> None:
+        super().__init__()
+        # SuperPoint uses 1 channel images
+        self.input_dim = 1
+
+        conv_blocks = []
+        conv_blocks.append(
+            SuperPointConvBlock(config, self.input_dim, config.encoder_hidden_sizes[0], add_pooling=True)
+        )
+        for i in range(1, len(config.encoder_hidden_sizes) - 1):
+            conv_blocks.append(
+                SuperPointConvBlock(
+                    config, config.encoder_hidden_sizes[i - 1], config.encoder_hidden_sizes[i], add_pooling=True
+                )
+            )
+        conv_blocks.append(
+            SuperPointConvBlock(
+                config, config.encoder_hidden_sizes[-2], config.encoder_hidden_sizes[-1], add_pooling=False
+            )
+        )
+        self.conv_blocks = nn.ModuleList(conv_blocks)
+
+    def forward(
+        self,
+        input,
+        output_hidden_states: bool | None = False,
+        return_dict: bool | None = True,
+    ) -> tuple | BaseModelOutputWithNoAttention:
+        all_hidden_states = () if output_hidden_states else None
+
+        for conv_block in self.conv_blocks:
+            input = conv_block(input)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (input,)
+        output = input
+        if not return_dict:
+            return tuple(v for v in [output, all_hidden_states] if v is not None)
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=output,
+            hidden_states=all_hidden_states,
+        )
+
+
+class SuperPointInterestPointDecoder(nn.Module):
+    """
+    The SuperPointInterestPointDecoder uses the output of the SuperPointEncoder to compute the keypoint with scores.
+    The scores are first computed by a convolutional layer, then a softmax is applied to get a probability distribution
+    over the 65 possible keypoint classes. The keypoints are then extracted from the scores by thresholding and
+    non-maximum suppression. Post-processing is then applied to remove keypoints too close to the image borders as well
+    as to keep only the k keypoints with highest score.
+    """
+
+    def __init__(self, config: SuperPointConfig) -> None:
+        super().__init__()
+        self.keypoint_threshold = config.keypoint_threshold
+        self.max_keypoints = config.max_keypoints
+        self.nms_radius = config.nms_radius
+        self.border_removal_distance = config.border_removal_distance
+
+        self.relu = nn.ReLU(inplace=True)
+        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.conv_score_a = nn.Conv2d(
+            config.encoder_hidden_sizes[-1],
+            config.decoder_hidden_size,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+        self.conv_score_b = nn.Conv2d(
+            config.decoder_hidden_size, config.keypoint_decoder_dim, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, encoded: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        scores = self._get_pixel_scores(encoded)
+        keypoints, scores = self._extract_keypoints(scores)
+
+        return keypoints, scores
+
+    def _get_pixel_scores(self, encoded: torch.Tensor) -> torch.Tensor:
+        """Based on the encoder output, compute the scores for each pixel of the image"""
+        scores = self.relu(self.conv_score_a(encoded))
+        scores = self.conv_score_b(scores)
+        scores = nn.functional.softmax(scores, 1)[:, :-1]
+        batch_size, _, height, width = scores.shape
+        scores = scores.permute(0, 2, 3, 1).reshape(batch_size, height, width, 8, 8)
+        scores = scores.permute(0, 1, 3, 2, 4).reshape(batch_size, height * 8, width * 8)
+        scores = simple_nms(scores, self.nms_radius)
+        return scores
+
+    def _extract_keypoints(self, scores: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        Based on their scores, extract the pixels that represent the keypoints that will be used for descriptors computation.
+        The keypoints are in the form of relative (x, y) coordinates.
+        """
+        _, height, width = scores.shape
+
+        # Threshold keypoints by score value
+        keypoints = torch.nonzero(scores[0] > self.keypoint_threshold)
+        scores = scores[0][tuple(keypoints.t())]
+
+        # Discard keypoints near the image borders
+        keypoints, scores = remove_keypoints_from_borders(
+            keypoints, scores, self.border_removal_distance, height * 8, width * 8
+        )
+
+        # Keep the k keypoints with highest score
+        if self.max_keypoints >= 0:
+            keypoints, scores = top_k_keypoints(keypoints, scores, self.max_keypoints)
+
+        # Convert (y, x) to (x, y)
+        keypoints = torch.flip(keypoints, [1]).to(scores.dtype)
+
+        return keypoints, scores
+
+
+class SuperPointDescriptorDecoder(nn.Module):
+    """
+    The SuperPointDescriptorDecoder uses the outputs of both the SuperPointEncoder and the
+    SuperPointInterestPointDecoder to compute the descriptors at the keypoints locations.
+
+    The descriptors are first computed by a convolutional layer, then normalized to have a norm of 1. The descriptors
+    are then interpolated at the keypoints locations.
+    """
+
+    def __init__(self, config: SuperPointConfig) -> None:
+        super().__init__()
+
+        self.relu = nn.ReLU(inplace=True)
+        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.conv_descriptor_a = nn.Conv2d(
+            config.encoder_hidden_sizes[-1],
+            config.decoder_hidden_size,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+        self.conv_descriptor_b = nn.Conv2d(
+            config.decoder_hidden_size,
+            config.descriptor_decoder_dim,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        )
+
+    def forward(self, encoded: torch.Tensor, keypoints: torch.Tensor) -> torch.Tensor:
+        """Based on the encoder output and the keypoints, compute the descriptors for each keypoint"""
+        descriptors = self.conv_descriptor_b(self.relu(self.conv_descriptor_a(encoded)))
+        descriptors = nn.functional.normalize(descriptors, p=2, dim=1)
+
+        descriptors = self._sample_descriptors(keypoints[None], descriptors[0][None], 8)[0]
+
+        # [descriptor_dim, num_keypoints] -> [num_keypoints, descriptor_dim]
+        descriptors = torch.transpose(descriptors, 0, 1)
+
+        return descriptors
+
+    @staticmethod
+    def _sample_descriptors(keypoints, descriptors, scale: int = 8) -> torch.Tensor:
+        """Interpolate descriptors at keypoint locations"""
+        batch_size, num_channels, height, width = descriptors.shape
+        keypoints = keypoints - scale / 2 + 0.5
+        divisor = torch.tensor([[(width * scale - scale / 2 - 0.5), (height * scale - scale / 2 - 0.5)]])
+        divisor = divisor.to(keypoints)
+        keypoints /= divisor
+        keypoints = keypoints * 2 - 1  # normalize to (-1, 1)
+        kwargs = {"align_corners": True}
+        # [batch_size, num_channels, num_keypoints, 2] -> [batch_size, num_channels, num_keypoints, 2]
+        keypoints = keypoints.view(batch_size, 1, -1, 2)
+        descriptors = nn.functional.grid_sample(descriptors, keypoints, mode="bilinear", **kwargs)
+        # [batch_size, descriptor_decoder_dim, num_channels, num_keypoints] -> [batch_size, descriptor_decoder_dim, num_keypoints]
+        descriptors = descriptors.reshape(batch_size, num_channels, -1)
+        descriptors = nn.functional.normalize(descriptors, p=2, dim=1)
+        return descriptors
+
+
+@auto_docstring
+class SuperPointPreTrainedModel(PreTrainedModel):
+    config: SuperPointConfig
+    base_model_prefix = "superpoint"
+    main_input_name = "pixel_values"
+    input_modalities = ("image",)
+    supports_gradient_checkpointing = False
+
+    def extract_one_channel_pixel_values(self, pixel_values: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        Assuming pixel_values has shape (batch_size, 3, height, width), and that all channels values are the same,
+        extract the first channel value to get a tensor of shape (batch_size, 1, height, width) for SuperPoint. This is
+        a workaround for the issue discussed in :
+        https://github.com/huggingface/transformers/pull/25786#issuecomment-1730176446
+
+        Args:
+            pixel_values: torch.FloatTensor of shape (batch_size, 3, height, width)
+
+        Returns:
+            pixel_values: torch.FloatTensor of shape (batch_size, 1, height, width)
+
+        """
+        return pixel_values[:, 0, :, :][:, None, :, :]
+
+
+@auto_docstring(
+    custom_intro="""
+    SuperPoint model outputting keypoints and descriptors.
+    """
+)
+class SuperPointForKeypointDetection(SuperPointPreTrainedModel):
+    """
+    SuperPoint model. It consists of a SuperPointEncoder, a SuperPointInterestPointDecoder and a
+    SuperPointDescriptorDecoder. SuperPoint was proposed in `SuperPoint: Self-Supervised Interest Point Detection and
+    Description <https://huggingface.co/papers/1712.07629>`__ by Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. It
+    is a fully convolutional neural network that extracts keypoints and descriptors from an image. It is trained in a
+    self-supervised manner, using a combination of a photometric loss and a loss based on the homographic adaptation of
+    keypoints. It is made of a convolutional encoder and two decoders: one for keypoints and one for descriptors.
+    """
+
+    def __init__(self, config: SuperPointConfig) -> None:
+        super().__init__(config)
+
+        self.config = config
+
+        self.encoder = SuperPointEncoder(config)
+        self.keypoint_decoder = SuperPointInterestPointDecoder(config)
+        self.descriptor_decoder = SuperPointDescriptorDecoder(config)
+
+        self.post_init()
+
+    @auto_docstring
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        labels: torch.LongTensor | None = None,
+        output_hidden_states: bool | None = None,
+        return_dict: bool | None = None,
+        **kwargs,
+    ) -> tuple | SuperPointKeypointDescriptionOutput:
+        r"""
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, SuperPointForKeypointDetection
+        >>> import torch
+        >>> from PIL import Image
+        >>> import httpx
+        >>> from io import BytesIO
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
+
+        >>> processor = AutoImageProcessor.from_pretrained("magic-leap-community/superpoint")
+        >>> model = SuperPointForKeypointDetection.from_pretrained("magic-leap-community/superpoint")
+
+        >>> inputs = processor(image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        ```"""
+        loss = None
+        if labels is not None:
+            raise ValueError("SuperPoint does not support training for now.")
+
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        pixel_values = self.extract_one_channel_pixel_values(pixel_values)
+
+        batch_size, _, height, width = pixel_values.shape
+
+        encoder_outputs = self.encoder(
+            pixel_values,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+
+        list_keypoints_scores = [
+            self.keypoint_decoder(last_hidden_state[None, ...]) for last_hidden_state in last_hidden_state
+        ]
+
+        list_keypoints = [keypoints_scores[0] for keypoints_scores in list_keypoints_scores]
+        list_scores = [keypoints_scores[1] for keypoints_scores in list_keypoints_scores]
+
+        list_descriptors = [
+            self.descriptor_decoder(last_hidden_state[None, ...], keypoints[None, ...])
+            for last_hidden_state, keypoints in zip(last_hidden_state, list_keypoints)
+        ]
+
+        maximum_num_keypoints = max(keypoints.shape[0] for keypoints in list_keypoints)
+
+        keypoints = torch.zeros((batch_size, maximum_num_keypoints, 2), device=pixel_values.device)
+        scores = torch.zeros((batch_size, maximum_num_keypoints), device=pixel_values.device)
+        descriptors = torch.zeros(
+            (batch_size, maximum_num_keypoints, self.config.descriptor_decoder_dim),
+            device=pixel_values.device,
+        )
+        mask = torch.zeros((batch_size, maximum_num_keypoints), device=pixel_values.device, dtype=torch.int)
+
+        for i, (_keypoints, _scores, _descriptors) in enumerate(zip(list_keypoints, list_scores, list_descriptors)):
+            keypoints[i, : _keypoints.shape[0]] = _keypoints
+            scores[i, : _scores.shape[0]] = _scores
+            descriptors[i, : _descriptors.shape[0]] = _descriptors
+            mask[i, : _scores.shape[0]] = 1
+
+        # Convert to relative coordinates
+        keypoints = keypoints / torch.tensor([width, height], device=keypoints.device)
+
+        hidden_states = encoder_outputs[1] if output_hidden_states else None
+        if not return_dict:
+            return tuple(v for v in [loss, keypoints, scores, descriptors, mask, hidden_states] if v is not None)
+
+        return SuperPointKeypointDescriptionOutput(
+            loss=loss,
+            keypoints=keypoints,
+            scores=scores,
+            descriptors=descriptors,
+            mask=mask,
+            hidden_states=hidden_states,
+        )
+
+
+__all__ = ["SuperPointForKeypointDetection", "SuperPointPreTrainedModel"]

LTA_openwebtext_dualt/mini_owt_logdirichlet/runs/owt_t5_elftokenized_full_len1024_C1_to_1024_pow1_d768_l12_h12_gbs512_8gpu_50ep_lr6e4_ema0p9999_elfopt_t5embed_unfixed_norm_stateprobadd_selfcond_ce_fast_20260606_144245/step_049000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2226a2bd8e48428f4db4ace5cf66e7599ac8f1fddc1dab9cb1b3d02c26d1b0b
+size 1031030114

LTA_openwebtext_dualt/mini_owt_logdirichlet/runs/owt_t5_elftokenized_full_len1024_C1_to_1024_pow1_d768_l12_h12_gbs512_8gpu_50ep_lr6e4_ema0p9999_elfopt_t5embed_unfixed_norm_stateprobadd_selfcond_ce_fast_20260606_144245/step_050000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e78dbec6f6736c9b082d50541356c64ca9ebecdfcd44a804fec78214b04b4538
+size 1031030114

LTA_openwebtext_dualt/mini_owt_logdirichlet/runs/owt_t5_elftokenized_full_len1024_C1_to_1024_pow1_d768_l12_h12_gbs512_8gpu_50ep_lr6e4_ema0p9999_elfopt_t5embed_unfixed_norm_stateprobadd_selfcond_ce_fast_20260606_144245/step_262000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9a2c4435e123f5672449243a3513dca3018c2789107eef4eae0637b52058e05
+size 1031030114

LTA_openwebtext_dualt/mini_owt_logdirichlet/runs/owt_t5_elftokenized_full_len1024_C1_to_1024_pow1_d768_l12_h12_gbs512_8gpu_50ep_lr6e4_ema0p9999_elfopt_t5embed_unfixed_norm_stateprobadd_selfcond_ce_fast_20260606_144245/step_295000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55e15c16ae46b6b5b03ffc62b504002d23cdde13fcc009402a48e552475659d9
+size 1031030114