Add files using upload-large-folder tool

dataset_inputs/training/llama3b_code_sft_data/READY.json

@@ -0,0 +1,13 @@
+{
+  "filename": "llama3b_code_sft_data.jsonl",
+  "kind": "coding_sft_dataset",
+  "manifest_filename": "llama3b_code_sft_data_manifest.json",
+  "manifest_sha256": "c746a11b5b02ac298775340819de1a29a5e34052bf9d183181db9de11eb855de",
+  "max_step": 200,
+  "min_step": 1,
+  "repo_id": "RL-Forgetting-Exp-2/llama3b_code_sft_data",
+  "revision": "185fa951b03ae71771118876d2faf37f1fb1eea2",
+  "rows": 16421,
+  "schema_version": 1,
+  "sha256": "69ec8901893709bcba86b78d434c5995f523c783916edffa0129aed8649040de"
+}

dataset_inputs/training/llama3b_code_sft_data/llama3b_code_sft_data_manifest.json

@@ -0,0 +1,33 @@
+{
+  "created": "2026-09-13 14:28:07",
+  "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/llama3p2_3b_it_mbpp_nobuf/rollouts/training",
+  "sha256": "69ec8901893709bcba86b78d434c5995f523c783916edffa0129aed8649040de",
+  "args": {
+    "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/llama3p2_3b_it_mbpp_nobuf/rollouts/training",
+    "output": "data/llama3b_code_sft_data.jsonl",
+    "score_threshold": 0.0,
+    "drop_saturated": true,
+    "min_step": 0,
+    "max_step": 0,
+    "num_rl_steps": 200,
+    "max_records": 0,
+    "audit": true,
+    "verbose_every": 0
+  },
+  "stats": {
+    "rollouts": 102400,
+    "saturated_groups": 6503,
+    "positives": 68445,
+    "dropped_saturated": 52024,
+    "kept": 16421
+  },
+  "n_examples": 16421,
+  "n_rl_steps": 200,
+  "step_first": 1,
+  "step_last": 200,
+  "order": "ascending step, then original line order within the step dump",
+  "audit": {
+    "passed": true,
+    "detail": "exact order match over 16,421 records"
+  }
+}

dataset_inputs/training/qwen3_1p7b_s500_code_sft_data/READY.json

@@ -0,0 +1,13 @@
+{
+  "filename": "qwen3_1p7b_s500_code_sft_data.jsonl",
+  "kind": "coding_sft_dataset",
+  "manifest_filename": "qwen3_1p7b_s500_code_sft_data_manifest.json",
+  "manifest_sha256": "7e509081c6051b9a58de0db6579a51f5c480fb73dc7129f201509eabf71ff65f",
+  "max_step": 500,
+  "min_step": 1,
+  "repo_id": "RL-Forgetting-Exp-2/qwen3_1p7b_s500_code_sft_data",
+  "revision": "40d77cf622eee6fb504f0ba7143942601c3766e3",
+  "rows": 25926,
+  "schema_version": 1,
+  "sha256": "08974359f21d570e60fe801ee98c139f5bd943d0c8411ec03bec40b4ca2a96a6"
+}

dataset_inputs/training/qwen3_1p7b_s500_code_sft_data/qwen3_1p7b_s500_code_sft_data_manifest.json

@@ -0,0 +1,33 @@
+{
+  "created": "2026-09-15 09:51:23",
+  "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/qwen3_1p7b_base_mbpp_nobuf/rollouts/training",
+  "sha256": "08974359f21d570e60fe801ee98c139f5bd943d0c8411ec03bec40b4ca2a96a6",
+  "args": {
+    "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/qwen3_1p7b_base_mbpp_nobuf/rollouts/training",
+    "output": "data/qwen3_1p7b_s500_code_sft_data.jsonl",
+    "score_threshold": 0.0,
+    "drop_saturated": true,
+    "min_step": 0,
+    "max_step": 0,
+    "num_rl_steps": 500,
+    "max_records": 0,
+    "audit": true,
+    "verbose_every": 0
+  },
+  "stats": {
+    "rollouts": 256000,
+    "saturated_groups": 20355,
+    "positives": 188766,
+    "dropped_saturated": 162840,
+    "kept": 25926
+  },
+  "n_examples": 25926,
+  "n_rl_steps": 500,
+  "step_first": 1,
+  "step_last": 500,
+  "order": "ascending step, then original line order within the step dump",
+  "audit": {
+    "passed": true,
+    "detail": "exact order match over 25,926 records"
+  }
+}

dataset_inputs/training/qwen3b_code_sft_data_s300/READY.json

@@ -0,0 +1,13 @@
+{
+  "filename": "qwen3b_code_sft_data_s300.jsonl",
+  "kind": "coding_sft_dataset",
+  "manifest_filename": "qwen3b_code_sft_data_s300_manifest.json",
+  "manifest_sha256": "a885aed357462f4a4fefb11e06e6fd8c881811597d6b0f9ba739dcc87ced67ac",
+  "max_step": 300,
+  "min_step": 1,
+  "repo_id": "RL-Forgetting-Exp-2/qwen3b_code_sft_data_s300",
+  "revision": "c9f48cd5ac34894be0b38c35aebba931a022bada",
+  "rows": 25867,
+  "schema_version": 1,
+  "sha256": "22510155770eeff3f05ce0a5d1286f6f051b6baf3480a1276d7dc9da8230d345"
+}

dataset_inputs/training/qwen3b_code_sft_data_s300/qwen3b_code_sft_data_s300_manifest.json

@@ -0,0 +1,33 @@
+{
+  "created": "2026-09-22 17:41:17",
+  "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/qwen3b_mbpp_nobuf/rollouts/training",
+  "sha256": "22510155770eeff3f05ce0a5d1286f6f051b6baf3480a1276d7dc9da8230d345",
+  "args": {
+    "rollout_dir": "/scratch/js15262/rl-forgetting-experiments/results/code/qwen3b_mbpp_nobuf/rollouts/training",
+    "output": "data/qwen3b_code_sft_data_s300.jsonl",
+    "score_threshold": 0.0,
+    "drop_saturated": true,
+    "min_step": 0,
+    "max_step": 300,
+    "num_rl_steps": 0,
+    "max_records": 0,
+    "audit": true,
+    "verbose_every": 200
+  },
+  "stats": {
+    "rollouts": 153600,
+    "saturated_groups": 10643,
+    "positives": 111011,
+    "dropped_saturated": 85144,
+    "kept": 25867
+  },
+  "n_examples": 25867,
+  "n_rl_steps": 300,
+  "step_first": 1,
+  "step_last": 300,
+  "order": "ascending step, then original line order within the step dump",
+  "audit": {
+    "passed": true,
+    "detail": "exact order match over 25,867 records"
+  }
+}

training/llama3b_sft_rp_ce_lam1/metrics/train_metrics.jsonl

@@ -0,0 +1,65 @@
+{"global_step": 1, "epoch": 0, "batch_id": 63, "num_supervised_tokens": 97674, "learning_rate": 1e-05, "batch_loss": 1.3036121778480152, "token_entropy": 1.3033737616757028, "gradient_norm_pre_clip": 1.6959883563993348, "gradient_norm_post_clip": 1.6959883563993348, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.1953125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.1953125, "max_sequence_length": 3030.0, "replay_loss": NaN, "replay_num_supervised_tokens": 0, "total_loss": 1.3036121778480152, "replay_lambda": 1.0, "buffer_rows": 64, "buffer_prompts": 54}
+{"global_step": 2, "epoch": 0, "batch_id": 127, "num_supervised_tokens": 92723, "learning_rate": 1e-05, "batch_loss": 0.6894348270866253, "token_entropy": 0.6530921676807052, "gradient_norm_pre_clip": 2.7457620022793843, "gradient_norm_post_clip": 2.7457620022793843, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.666015625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.666015625, "max_sequence_length": 3030.0, "replay_loss": 8.335680656527433, "replay_num_supervised_tokens": 434615, "total_loss": 9.025115483614059, "replay_lambda": 1.0, "buffer_rows": 128, "buffer_prompts": 107, "replay_draw_n_groups": 64, "replay_draw_mean_ce": 8.63937434175882, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 3, "epoch": 0, "batch_id": 191, "num_supervised_tokens": 92641, "learning_rate": 1e-05, "batch_loss": 0.7003915629167659, "token_entropy": 0.7472277903481933, "gradient_norm_pre_clip": 4.497192828773844, "gradient_norm_post_clip": 4.497192828773844, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.666015625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.666015625, "max_sequence_length": 3030.0, "replay_loss": 8.0839933295982, "replay_num_supervised_tokens": 422629, "total_loss": 8.784384892514966, "replay_lambda": 1.0, "buffer_rows": 192, "buffer_prompts": 127, "replay_draw_n_groups": 128, "replay_draw_mean_ce": 8.625393401954977, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 4, "epoch": 0, "batch_id": 255, "num_supervised_tokens": 87718, "learning_rate": 1e-05, "batch_loss": 0.9031168897304659, "token_entropy": 0.9235762928684579, "gradient_norm_pre_clip": 1.9035836004212159, "gradient_norm_post_clip": 1.9035836004212159, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.666015625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.666015625, "max_sequence_length": 3030.0, "replay_loss": 7.886701050826728, "replay_num_supervised_tokens": 380195, "total_loss": 8.789817940557194, "replay_lambda": 1.0, "buffer_rows": 256, "buffer_prompts": 148, "replay_draw_n_groups": 192, "replay_draw_mean_ce": 8.542578433550553, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 5, "epoch": 0, "batch_id": 319, "num_supervised_tokens": 82516, "learning_rate": 1e-05, "batch_loss": 0.8565721539161542, "token_entropy": 0.7765942675992575, "gradient_norm_pre_clip": 2.6329424033768225, "gradient_norm_post_clip": 2.6329424033768225, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.666015625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.666015625, "max_sequence_length": 3030.0, "replay_loss": 8.080923703276843, "replay_num_supervised_tokens": 393799, "total_loss": 8.937495857192998, "replay_lambda": 1.0, "buffer_rows": 320, "buffer_prompts": 168, "replay_draw_n_groups": 256, "replay_draw_mean_ce": 8.493977607457873, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 6, "epoch": 0, "batch_id": 383, "num_supervised_tokens": 78503, "learning_rate": 1e-05, "batch_loss": 1.1520204392413527, "token_entropy": 1.0954276529611005, "gradient_norm_pre_clip": 1.978843775949324, "gradient_norm_post_clip": 1.978843775949324, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.666015625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.666015625, "max_sequence_length": 3030.0, "replay_loss": 8.302541697225156, "replay_num_supervised_tokens": 421855, "total_loss": 9.454562136466508, "replay_lambda": 1.0, "buffer_rows": 384, "buffer_prompts": 182, "replay_draw_n_groups": 320, "replay_draw_mean_ce": 8.317220309208794, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 7, "epoch": 0, "batch_id": 447, "num_supervised_tokens": 75606, "learning_rate": 1e-05, "batch_loss": 1.1976064150116819, "token_entropy": 1.183326557616445, "gradient_norm_pre_clip": 1.8241927002015876, "gradient_norm_post_clip": 1.8241927002015876, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.744140625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.744140625, "max_sequence_length": 3030.0, "replay_loss": 8.527549628385453, "replay_num_supervised_tokens": 470885, "total_loss": 9.725156043397135, "replay_lambda": 1.0, "buffer_rows": 448, "buffer_prompts": 190, "replay_draw_n_groups": 384, "replay_draw_mean_ce": 8.245614468141127, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 8, "epoch": 0, "batch_id": 511, "num_supervised_tokens": 65460, "learning_rate": 1e-05, "batch_loss": 0.6367896821890499, "token_entropy": 0.6493041185658167, "gradient_norm_pre_clip": 1.8959335335130396, "gradient_norm_post_clip": 1.8959335335130396, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 8.42786227623966, "replay_num_supervised_tokens": 460655, "total_loss": 9.06465195842871, "replay_lambda": 1.0, "buffer_rows": 512, "buffer_prompts": 197, "replay_draw_n_groups": 448, "replay_draw_mean_ce": 8.16910215903107, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 9, "epoch": 0, "batch_id": 575, "num_supervised_tokens": 60380, "learning_rate": 1e-05, "batch_loss": 0.6645281846049843, "token_entropy": 0.6811467412371792, "gradient_norm_pre_clip": 1.8926152417007243, "gradient_norm_post_clip": 1.8926152417007243, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 8.28472549643829, "replay_num_supervised_tokens": 470530, "total_loss": 8.949253681043274, "replay_lambda": 1.0, "buffer_rows": 576, "buffer_prompts": 200, "replay_draw_n_groups": 512, "replay_draw_mean_ce": 7.956049544990784, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 10, "epoch": 0, "batch_id": 639, "num_supervised_tokens": 60964, "learning_rate": 1e-05, "batch_loss": 0.5209809438427574, "token_entropy": 0.5336154938432444, "gradient_norm_pre_clip": 1.7689265579940463, "gradient_norm_post_clip": 1.7689265579940463, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 8.123257812373513, "replay_num_supervised_tokens": 425019, "total_loss": 8.64423875621627, "replay_lambda": 1.0, "buffer_rows": 640, "buffer_prompts": 204, "replay_draw_n_groups": 576, "replay_draw_mean_ce": 0.3964144206661891, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.407111505791, "parameter_l2_from_init": 2.6456406728592214, "relative_parameter_l2_from_init": 0.00238045348497299, "update_norm": 2.6456406728592214, "relative_update_norm": 0.002380442454857764}
+{"global_step": 11, "epoch": 0, "batch_id": 703, "num_supervised_tokens": 54290, "learning_rate": 1e-05, "batch_loss": 0.48360612405531644, "token_entropy": 0.47421949516235956, "gradient_norm_pre_clip": 2.244485384484945, "gradient_norm_post_clip": 2.244485384484945, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 8.135307800165606, "replay_num_supervised_tokens": 448941, "total_loss": 8.618913924220923, "replay_lambda": 1.0, "buffer_rows": 704, "buffer_prompts": 206, "replay_draw_n_groups": 640, "replay_draw_mean_ce": 7.724385161598209, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 12, "epoch": 0, "batch_id": 767, "num_supervised_tokens": 54744, "learning_rate": 1e-05, "batch_loss": 0.43163532276459426, "token_entropy": 0.42668402200202993, "gradient_norm_pre_clip": 2.925561019730532, "gradient_norm_post_clip": 2.925561019730532, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 7.913368445978619, "replay_num_supervised_tokens": 409372, "total_loss": 8.345003768743213, "replay_lambda": 1.0, "buffer_rows": 768, "buffer_prompts": 209, "replay_draw_n_groups": 704, "replay_draw_mean_ce": 7.56301167966693, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 13, "epoch": 0, "batch_id": 831, "num_supervised_tokens": 48778, "learning_rate": 1e-05, "batch_loss": 0.43828473494365366, "token_entropy": 0.4321255105375349, "gradient_norm_pre_clip": 5.249941878536121, "gradient_norm_post_clip": 5.249941878536121, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 7.707152831319, "replay_num_supervised_tokens": 403563, "total_loss": 8.145437566262654, "replay_lambda": 1.0, "buffer_rows": 832, "buffer_prompts": 212, "replay_draw_n_groups": 768, "replay_draw_mean_ce": 7.411626493384676, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 14, "epoch": 0, "batch_id": 895, "num_supervised_tokens": 48298, "learning_rate": 1e-05, "batch_loss": 0.4103400055739437, "token_entropy": 0.417402464517417, "gradient_norm_pre_clip": 7.260167908387989, "gradient_norm_post_clip": 7.260167908387989, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 7.578183646842082, "replay_num_supervised_tokens": 390961, "total_loss": 7.988523652416026, "replay_lambda": 1.0, "buffer_rows": 896, "buffer_prompts": 214, "replay_draw_n_groups": 832, "replay_draw_mean_ce": 7.237308801507372, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 15, "epoch": 0, "batch_id": 959, "num_supervised_tokens": 44942, "learning_rate": 1e-05, "batch_loss": 0.35589037957083136, "token_entropy": 0.3676269472342966, "gradient_norm_pre_clip": 3.5746551045538717, "gradient_norm_post_clip": 3.5746551045538717, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 7.425939946918765, "replay_num_supervised_tokens": 378692, "total_loss": 7.7818303264895965, "replay_lambda": 1.0, "buffer_rows": 960, "buffer_prompts": 216, "replay_draw_n_groups": 896, "replay_draw_mean_ce": 7.094080149091627, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 16, "epoch": 0, "batch_id": 1023, "num_supervised_tokens": 40349, "learning_rate": 1e-05, "batch_loss": 0.3409684640541331, "token_entropy": 0.3529405497675028, "gradient_norm_pre_clip": 9.100633531805924, "gradient_norm_post_clip": 9.100633531805924, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 7.210555704201064, "replay_num_supervised_tokens": 358536, "total_loss": 7.551524168255197, "replay_lambda": 1.0, "buffer_rows": 1024, "buffer_prompts": 218, "replay_draw_n_groups": 960, "replay_draw_mean_ce": 0.8512321437955437, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 17, "epoch": 0, "batch_id": 1087, "num_supervised_tokens": 41901, "learning_rate": 1e-05, "batch_loss": 0.3277391136780462, "token_entropy": 0.34149096210754953, "gradient_norm_pre_clip": 11.478637314779277, "gradient_norm_post_clip": 11.478637314779277, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.966162923671571, "replay_num_supervised_tokens": 333416, "total_loss": 7.293902037349617, "replay_lambda": 1.0, "buffer_rows": 1088, "buffer_prompts": 223, "replay_draw_n_groups": 1024, "replay_draw_mean_ce": 0.7152841567993165, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 18, "epoch": 0, "batch_id": 1151, "num_supervised_tokens": 37655, "learning_rate": 1e-05, "batch_loss": 0.34376660929608194, "token_entropy": 0.3524521612548018, "gradient_norm_pre_clip": 6.210512307214839, "gradient_norm_post_clip": 6.210512307214839, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.681832901306593, "replay_num_supervised_tokens": 310612, "total_loss": 7.0255995106026745, "replay_lambda": 1.0, "buffer_rows": 1152, "buffer_prompts": 225, "replay_draw_n_groups": 1088, "replay_draw_mean_ce": 6.399287556773301, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 19, "epoch": 0, "batch_id": 1215, "num_supervised_tokens": 38084, "learning_rate": 1e-05, "batch_loss": 0.287988393195893, "token_entropy": 0.284267149831978, "gradient_norm_pre_clip": 23.22691861827933, "gradient_norm_post_clip": 23.22691861827933, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.621407393769446, "replay_num_supervised_tokens": 306596, "total_loss": 6.909395786965339, "replay_lambda": 1.0, "buffer_rows": 1216, "buffer_prompts": 226, "replay_draw_n_groups": 1152, "replay_draw_mean_ce": 5.507357670062155, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 20, "epoch": 0, "batch_id": 1279, "num_supervised_tokens": 36949, "learning_rate": 1e-05, "batch_loss": 0.28172141613096485, "token_entropy": 0.28515216609861194, "gradient_norm_pre_clip": 8.92815020136883, "gradient_norm_post_clip": 8.92815020136883, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.5368220643262, "replay_num_supervised_tokens": 309360, "total_loss": 6.818543480457165, "replay_lambda": 1.0, "buffer_rows": 1280, "buffer_prompts": 229, "replay_draw_n_groups": 1216, "replay_draw_mean_ce": 5.907104170780674, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.4154488806969, "parameter_l2_from_init": 4.567179056102271, "relative_parameter_l2_from_init": 0.004109385455147489, "update_norm": 2.5524617267865994, "relative_update_norm": 0.0022965865098934658}
+{"global_step": 21, "epoch": 0, "batch_id": 1343, "num_supervised_tokens": 36603, "learning_rate": 1e-05, "batch_loss": 0.28049205240330677, "token_entropy": 0.28428244248793444, "gradient_norm_pre_clip": 10.672056459841006, "gradient_norm_post_clip": 10.672056459841006, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.1539046057349935, "replay_num_supervised_tokens": 280215, "total_loss": 6.434396658138301, "replay_lambda": 1.0, "buffer_rows": 1344, "buffer_prompts": 232, "replay_draw_n_groups": 1280, "replay_draw_mean_ce": 0.3017151093818772, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 22, "epoch": 0, "batch_id": 1407, "num_supervised_tokens": 36508, "learning_rate": 1e-05, "batch_loss": 0.2545877973529374, "token_entropy": 0.25588707063419347, "gradient_norm_pre_clip": 9.583575282621169, "gradient_norm_post_clip": 9.583575282621169, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 6.105795342665953, "replay_num_supervised_tokens": 298098, "total_loss": 6.36038314001889, "replay_lambda": 1.0, "buffer_rows": 1408, "buffer_prompts": 233, "replay_draw_n_groups": 1344, "replay_draw_mean_ce": 0.19530468311562205, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 23, "epoch": 0, "batch_id": 1471, "num_supervised_tokens": 34482, "learning_rate": 1e-05, "batch_loss": 0.27563347000269817, "token_entropy": 0.2724245521227242, "gradient_norm_pre_clip": 8.996859054625547, "gradient_norm_post_clip": 8.996859054625547, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 5.898150903324447, "replay_num_supervised_tokens": 254522, "total_loss": 6.173784373327145, "replay_lambda": 1.0, "buffer_rows": 1472, "buffer_prompts": 233, "replay_draw_n_groups": 1408, "replay_draw_mean_ce": 0.16172760653208537, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 24, "epoch": 0, "batch_id": 1535, "num_supervised_tokens": 34610, "learning_rate": 1e-05, "batch_loss": 0.2627824160701207, "token_entropy": 0.25929422756041864, "gradient_norm_pre_clip": 12.396415652938447, "gradient_norm_post_clip": 12.396415652938447, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 5.695268659819678, "replay_num_supervised_tokens": 256279, "total_loss": 5.958051075889799, "replay_lambda": 1.0, "buffer_rows": 1536, "buffer_prompts": 235, "replay_draw_n_groups": 1472, "replay_draw_mean_ce": 3.217566741756135, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 25, "epoch": 0, "batch_id": 1599, "num_supervised_tokens": 36993, "learning_rate": 1e-05, "batch_loss": 0.24064380175032274, "token_entropy": 0.23551484650559878, "gradient_norm_pre_clip": 7.044237967928264, "gradient_norm_post_clip": 7.044237967928264, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 5.477628085505984, "replay_num_supervised_tokens": 241910, "total_loss": 5.718271887256306, "replay_lambda": 1.0, "buffer_rows": 1600, "buffer_prompts": 235, "replay_draw_n_groups": 1536, "replay_draw_mean_ce": 0.16041554245993356, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 26, "epoch": 0, "batch_id": 1663, "num_supervised_tokens": 35574, "learning_rate": 1e-05, "batch_loss": 0.2402539084349591, "token_entropy": 0.23952982216183744, "gradient_norm_pre_clip": 9.827983867859158, "gradient_norm_post_clip": 9.827983867859158, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 5.120559296789044, "replay_num_supervised_tokens": 218722, "total_loss": 5.360813205224003, "replay_lambda": 1.0, "buffer_rows": 1664, "buffer_prompts": 236, "replay_draw_n_groups": 1600, "replay_draw_mean_ce": 0.5721170150816208, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 27, "epoch": 0, "batch_id": 1727, "num_supervised_tokens": 31824, "learning_rate": 1e-05, "batch_loss": 0.239359658701011, "token_entropy": 0.24166771603205983, "gradient_norm_pre_clip": 6.644871616657087, "gradient_norm_post_clip": 6.644871616657087, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 4.648402800882456, "replay_num_supervised_tokens": 173479, "total_loss": 4.887762459583468, "replay_lambda": 1.0, "buffer_rows": 1728, "buffer_prompts": 237, "replay_draw_n_groups": 1664, "replay_draw_mean_ce": 0.38755693008650594, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 28, "epoch": 0, "batch_id": 1791, "num_supervised_tokens": 33894, "learning_rate": 1e-05, "batch_loss": 0.228190961581126, "token_entropy": 0.23769945183032143, "gradient_norm_pre_clip": 9.944736996565968, "gradient_norm_post_clip": 9.944736996565968, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 4.472608818690773, "replay_num_supervised_tokens": 187997, "total_loss": 4.7007997802718995, "replay_lambda": 1.0, "buffer_rows": 1792, "buffer_prompts": 242, "replay_draw_n_groups": 1728, "replay_draw_mean_ce": 4.846635445441989, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 29, "epoch": 0, "batch_id": 1855, "num_supervised_tokens": 32475, "learning_rate": 1e-05, "batch_loss": 0.2245268974990639, "token_entropy": 0.23767186077490873, "gradient_norm_pre_clip": 7.8643161030369475, "gradient_norm_post_clip": 7.8643161030369475, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 3.8428195666206446, "replay_num_supervised_tokens": 143475, "total_loss": 4.067346464119709, "replay_lambda": 1.0, "buffer_rows": 1856, "buffer_prompts": 243, "replay_draw_n_groups": 1792, "replay_draw_mean_ce": 0.4571836459561714, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 30, "epoch": 0, "batch_id": 1919, "num_supervised_tokens": 30711, "learning_rate": 1e-05, "batch_loss": 0.22749141131105502, "token_entropy": 0.24606460658734827, "gradient_norm_pre_clip": 9.058716125525159, "gradient_norm_post_clip": 9.058716125525159, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 3.445540690149017, "replay_num_supervised_tokens": 129162, "total_loss": 3.673032101460072, "replay_lambda": 1.0, "buffer_rows": 1920, "buffer_prompts": 243, "replay_draw_n_groups": 1856, "replay_draw_mean_ce": 0.2164423356362439, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.4254994240318, "parameter_l2_from_init": 6.347735347102114, "relative_parameter_l2_from_init": 0.0057114667474344166, "update_norm": 2.5842457809656727, "relative_update_norm": 0.002325163299118918}
+{"global_step": 31, "epoch": 0, "batch_id": 1983, "num_supervised_tokens": 31193, "learning_rate": 1e-05, "batch_loss": 0.19464887440055575, "token_entropy": 0.2179863352860109, "gradient_norm_pre_clip": 12.02999272663547, "gradient_norm_post_clip": 12.02999272663547, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 3.3132333765897717, "replay_num_supervised_tokens": 128989, "total_loss": 3.5078822509903276, "replay_lambda": 1.0, "buffer_rows": 1984, "buffer_prompts": 245, "replay_draw_n_groups": 1920, "replay_draw_mean_ce": 2.918063748267389, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 32, "epoch": 0, "batch_id": 2047, "num_supervised_tokens": 29719, "learning_rate": 1e-05, "batch_loss": 0.19464879197686896, "token_entropy": 0.22047884194148717, "gradient_norm_pre_clip": 11.036677920816537, "gradient_norm_post_clip": 11.036677920816537, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 2.7235363513666155, "replay_num_supervised_tokens": 105263, "total_loss": 2.9181851433434844, "replay_lambda": 1.0, "buffer_rows": 2048, "buffer_prompts": 246, "replay_draw_n_groups": 1984, "replay_draw_mean_ce": 0.2050583924298511, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 33, "epoch": 0, "batch_id": 2111, "num_supervised_tokens": 30714, "learning_rate": 1e-05, "batch_loss": 0.18642509300554044, "token_entropy": 0.2064733699933466, "gradient_norm_pre_clip": 9.043206299994004, "gradient_norm_post_clip": 9.043206299994004, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 2.398883753588881, "replay_num_supervised_tokens": 93844, "total_loss": 2.5853088465944216, "replay_lambda": 1.0, "buffer_rows": 2112, "buffer_prompts": 247, "replay_draw_n_groups": 2048, "replay_draw_mean_ce": 0.13281778494517008, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 34, "epoch": 0, "batch_id": 2175, "num_supervised_tokens": 29442, "learning_rate": 1e-05, "batch_loss": 0.15929148969041265, "token_entropy": 0.18383935576312513, "gradient_norm_pre_clip": 6.7539444245480516, "gradient_norm_post_clip": 6.7539444245480516, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 2.054853168003248, "replay_num_supervised_tokens": 83050, "total_loss": 2.2141446576936605, "replay_lambda": 1.0, "buffer_rows": 2176, "buffer_prompts": 248, "replay_draw_n_groups": 2112, "replay_draw_mean_ce": 0.44913447986949573, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 35, "epoch": 0, "batch_id": 2239, "num_supervised_tokens": 34180, "learning_rate": 1e-05, "batch_loss": 0.14537330868443057, "token_entropy": 0.17041582377630762, "gradient_norm_pre_clip": 7.256173714661079, "gradient_norm_post_clip": 7.256173714661079, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 1.478220270661372, "replay_num_supervised_tokens": 68531, "total_loss": 1.6235935793458027, "replay_lambda": 1.0, "buffer_rows": 2240, "buffer_prompts": 250, "replay_draw_n_groups": 2176, "replay_draw_mean_ce": 0.18608187737865983, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 36, "epoch": 0, "batch_id": 2303, "num_supervised_tokens": 30222, "learning_rate": 1e-05, "batch_loss": 0.15353787893762336, "token_entropy": 0.17751217721496423, "gradient_norm_pre_clip": 6.14456499938851, "gradient_norm_post_clip": 6.14456499938851, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 1.7120624594387892, "replay_num_supervised_tokens": 81976, "total_loss": 1.8656003383764126, "replay_lambda": 1.0, "buffer_rows": 2304, "buffer_prompts": 254, "replay_draw_n_groups": 2240, "replay_draw_mean_ce": 0.5178396379625475, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 37, "epoch": 0, "batch_id": 2367, "num_supervised_tokens": 31163, "learning_rate": 1e-05, "batch_loss": 0.14737689007952634, "token_entropy": 0.16940043050066228, "gradient_norm_pre_clip": 7.696314872643681, "gradient_norm_post_clip": 7.696314872643681, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 1.5179157955496383, "replay_num_supervised_tokens": 76106, "total_loss": 1.6652926856291645, "replay_lambda": 1.0, "buffer_rows": 2368, "buffer_prompts": 256, "replay_draw_n_groups": 2304, "replay_draw_mean_ce": 2.5387512960551697, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 38, "epoch": 0, "batch_id": 2431, "num_supervised_tokens": 28931, "learning_rate": 1e-05, "batch_loss": 0.14936654943121563, "token_entropy": 0.1740556207760828, "gradient_norm_pre_clip": 8.038916202181667, "gradient_norm_post_clip": 8.038916202181667, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2916.0, "replay_loss": 1.1493363587020022, "replay_num_supervised_tokens": 70294, "total_loss": 1.298702908133218, "replay_lambda": 1.0, "buffer_rows": 2432, "buffer_prompts": 256, "replay_draw_n_groups": 2368, "replay_draw_mean_ce": 0.2098518830758554, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 39, "epoch": 0, "batch_id": 2495, "num_supervised_tokens": 29457, "learning_rate": 1e-05, "batch_loss": 0.12923849878387098, "token_entropy": 0.15487326475576546, "gradient_norm_pre_clip": 4.8227912853085595, "gradient_norm_post_clip": 4.8227912853085595, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.8179018737113072, "replay_num_supervised_tokens": 56226, "total_loss": 0.9471403724951781, "replay_lambda": 1.0, "buffer_rows": 2496, "buffer_prompts": 257, "replay_draw_n_groups": 2432, "replay_draw_mean_ce": 0.13743006216513143, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 40, "epoch": 0, "batch_id": 2559, "num_supervised_tokens": 30001, "learning_rate": 1e-05, "batch_loss": 0.1232631024569129, "token_entropy": 0.14675372335331538, "gradient_norm_pre_clip": 7.952752147397022, "gradient_norm_post_clip": 7.952752147397022, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.9675837088686653, "replay_num_supervised_tokens": 69712, "total_loss": 1.0908468113255783, "replay_lambda": 1.0, "buffer_rows": 2560, "buffer_prompts": 258, "replay_draw_n_groups": 2496, "replay_draw_mean_ce": 0.10797242264249432, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.433455084411, "parameter_l2_from_init": 7.890952706360191, "relative_parameter_l2_from_init": 0.0070999989009511615, "update_norm": 2.3928761633099582, "relative_update_norm": 0.0021529639515199084}
+{"global_step": 41, "epoch": 0, "batch_id": 2623, "num_supervised_tokens": 29105, "learning_rate": 1e-05, "batch_loss": 0.13414801958928535, "token_entropy": 0.1426922713454862, "gradient_norm_pre_clip": 6.614646150413235, "gradient_norm_post_clip": 6.614646150413235, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2916.0, "replay_loss": 0.7310973920397825, "replay_num_supervised_tokens": 56361, "total_loss": 0.8652454116290679, "replay_lambda": 1.0, "buffer_rows": 2624, "buffer_prompts": 258, "replay_draw_n_groups": 2560, "replay_draw_mean_ce": 0.319518676170936, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 42, "epoch": 0, "batch_id": 2687, "num_supervised_tokens": 28971, "learning_rate": 1e-05, "batch_loss": 0.10840614318673035, "token_entropy": 0.1284433276449773, "gradient_norm_pre_clip": 10.486239029274286, "gradient_norm_post_clip": 10.486239029274286, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.865206484999628, "replay_num_supervised_tokens": 63567, "total_loss": 0.9736126281863584, "replay_lambda": 1.0, "buffer_rows": 2688, "buffer_prompts": 259, "replay_draw_n_groups": 2624, "replay_draw_mean_ce": 0.4012954200854142, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 43, "epoch": 0, "batch_id": 2751, "num_supervised_tokens": 28111, "learning_rate": 1e-05, "batch_loss": 0.10769918861671292, "token_entropy": 0.12291809226010254, "gradient_norm_pre_clip": 9.11577082570239, "gradient_norm_post_clip": 9.11577082570239, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2916.0, "replay_loss": 0.651217383238228, "replay_num_supervised_tokens": 54691, "total_loss": 0.7589165718549409, "replay_lambda": 1.0, "buffer_rows": 2752, "buffer_prompts": 259, "replay_draw_n_groups": 2688, "replay_draw_mean_ce": 0.25334760877821183, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 44, "epoch": 0, "batch_id": 2815, "num_supervised_tokens": 29323, "learning_rate": 1e-05, "batch_loss": 0.10699571617966792, "token_entropy": 0.12182286183838872, "gradient_norm_pre_clip": 1.9576470857661232, "gradient_norm_post_clip": 1.9576470857661232, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1030.0, "replay_loss": 0.4709844593082299, "replay_num_supervised_tokens": 49817, "total_loss": 0.5779801754878978, "replay_lambda": 1.0, "buffer_rows": 2816, "buffer_prompts": 259, "replay_draw_n_groups": 2752, "replay_draw_mean_ce": 0.2313112031429186, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 45, "epoch": 0, "batch_id": 2879, "num_supervised_tokens": 26914, "learning_rate": 1e-05, "batch_loss": 0.10615110028374013, "token_entropy": 0.12396546133721292, "gradient_norm_pre_clip": 8.535142404127939, "gradient_norm_post_clip": 8.535142404127939, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2750.0, "replay_loss": 0.7206510688944783, "replay_num_supervised_tokens": 55818, "total_loss": 0.8268021691782184, "replay_lambda": 1.0, "buffer_rows": 2880, "buffer_prompts": 259, "replay_draw_n_groups": 2816, "replay_draw_mean_ce": 0.6710864735298416, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 46, "epoch": 0, "batch_id": 2943, "num_supervised_tokens": 25515, "learning_rate": 1e-05, "batch_loss": 0.1104128938496066, "token_entropy": 0.12513636563447045, "gradient_norm_pre_clip": 3.9942358157892377, "gradient_norm_post_clip": 3.9942358157892377, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.5756403757943591, "replay_num_supervised_tokens": 52422, "total_loss": 0.6860532696439657, "replay_lambda": 1.0, "buffer_rows": 2944, "buffer_prompts": 259, "replay_draw_n_groups": 2880, "replay_draw_mean_ce": 0.10046392605628496, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 47, "epoch": 0, "batch_id": 3007, "num_supervised_tokens": 30018, "learning_rate": 1e-05, "batch_loss": 0.08564601253578648, "token_entropy": 0.09959821804620271, "gradient_norm_pre_clip": 2.9570288223756878, "gradient_norm_post_clip": 2.9570288223756878, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2683.0, "replay_loss": 0.5255741983982676, "replay_num_supervised_tokens": 52274, "total_loss": 0.6112202109340541, "replay_lambda": 1.0, "buffer_rows": 3008, "buffer_prompts": 261, "replay_draw_n_groups": 2944, "replay_draw_mean_ce": 0.5874360765729632, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 48, "epoch": 0, "batch_id": 3071, "num_supervised_tokens": 30651, "learning_rate": 1e-05, "batch_loss": 0.08607116701279321, "token_entropy": 0.10474265745337887, "gradient_norm_pre_clip": 4.941739035162133, "gradient_norm_post_clip": 4.941739035162133, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2750.0, "replay_loss": 0.49790174539144016, "replay_num_supervised_tokens": 51435, "total_loss": 0.5839729124042334, "replay_lambda": 1.0, "buffer_rows": 3072, "buffer_prompts": 261, "replay_draw_n_groups": 3008, "replay_draw_mean_ce": 0.4046428285796067, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 49, "epoch": 0, "batch_id": 3135, "num_supervised_tokens": 28413, "learning_rate": 1e-05, "batch_loss": 0.11111577796607415, "token_entropy": 0.11918847314937102, "gradient_norm_pre_clip": 3.62705715393032, "gradient_norm_post_clip": 3.62705715393032, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.46605339128097256, "replay_num_supervised_tokens": 49332, "total_loss": 0.5771691692470468, "replay_lambda": 1.0, "buffer_rows": 3136, "buffer_prompts": 261, "replay_draw_n_groups": 3072, "replay_draw_mean_ce": 0.06849942888532366, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 50, "epoch": 0, "batch_id": 3199, "num_supervised_tokens": 27750, "learning_rate": 1e-05, "batch_loss": 0.0900980373131665, "token_entropy": 0.10819172988066802, "gradient_norm_pre_clip": 2.051618318034205, "gradient_norm_post_clip": 2.051618318034205, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 729.0, "replay_loss": 0.3723364453151571, "replay_num_supervised_tokens": 44372, "total_loss": 0.4624344826283236, "replay_lambda": 1.0, "buffer_rows": 3200, "buffer_prompts": 261, "replay_draw_n_groups": 3136, "replay_draw_mean_ce": 0.24089094068183273, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.4368802491322, "parameter_l2_from_init": 8.869197051761171, "relative_parameter_l2_from_init": 0.00798018840856416, "update_norm": 1.5938508775787, "relative_update_norm": 0.0014340453388782934}
+{"global_step": 51, "epoch": 0, "batch_id": 3263, "num_supervised_tokens": 24101, "learning_rate": 1e-05, "batch_loss": 0.09315593021784903, "token_entropy": 0.1035290941609015, "gradient_norm_pre_clip": 7.792418294731359, "gradient_norm_post_clip": 7.792418294731359, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2916.0, "replay_loss": 0.59197793091178, "replay_num_supervised_tokens": 54630, "total_loss": 0.685133861129629, "replay_lambda": 1.0, "buffer_rows": 3264, "buffer_prompts": 261, "replay_draw_n_groups": 3200, "replay_draw_mean_ce": 0.33649908529745565, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 52, "epoch": 0, "batch_id": 3327, "num_supervised_tokens": 29272, "learning_rate": 1e-05, "batch_loss": 0.09914495827800442, "token_entropy": 0.09568310266115783, "gradient_norm_pre_clip": 5.481024269256962, "gradient_norm_post_clip": 5.481024269256962, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.48116988282173945, "replay_num_supervised_tokens": 49781, "total_loss": 0.5803148410997438, "replay_lambda": 1.0, "buffer_rows": 3328, "buffer_prompts": 262, "replay_draw_n_groups": 3264, "replay_draw_mean_ce": 0.061790545090385104, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 53, "epoch": 0, "batch_id": 3391, "num_supervised_tokens": 29206, "learning_rate": 1e-05, "batch_loss": 0.117712059428278, "token_entropy": 0.09059335482918073, "gradient_norm_pre_clip": 3.6825435923883143, "gradient_norm_post_clip": 3.6825435923883143, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1030.0, "replay_loss": 0.3782027697143963, "replay_num_supervised_tokens": 41978, "total_loss": 0.4959148291426743, "replay_lambda": 1.0, "buffer_rows": 3392, "buffer_prompts": 263, "replay_draw_n_groups": 3328, "replay_draw_mean_ce": 0.32656135559082033, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 54, "epoch": 0, "batch_id": 3455, "num_supervised_tokens": 28302, "learning_rate": 1e-05, "batch_loss": 0.08911515506549497, "token_entropy": 0.11158868545576436, "gradient_norm_pre_clip": 3.31554326914073, "gradient_norm_post_clip": 3.31554326914073, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.4424611944706824, "replay_num_supervised_tokens": 47269, "total_loss": 0.5315763495361774, "replay_lambda": 1.0, "buffer_rows": 3456, "buffer_prompts": 263, "replay_draw_n_groups": 3392, "replay_draw_mean_ce": 0.14977268580972714, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 55, "epoch": 0, "batch_id": 3519, "num_supervised_tokens": 24971, "learning_rate": 1e-05, "batch_loss": 0.09504209631986614, "token_entropy": 0.12234752841097186, "gradient_norm_pre_clip": 4.545595464101638, "gradient_norm_post_clip": 4.545595464101638, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2607.0, "replay_loss": 0.46701081021075735, "replay_num_supervised_tokens": 47324, "total_loss": 0.5620529065306235, "replay_lambda": 1.0, "buffer_rows": 3520, "buffer_prompts": 264, "replay_draw_n_groups": 3456, "replay_draw_mean_ce": 0.16934591046931072, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 56, "epoch": 0, "batch_id": 3583, "num_supervised_tokens": 31094, "learning_rate": 1e-05, "batch_loss": 0.0745987084747461, "token_entropy": 0.0976024787663846, "gradient_norm_pre_clip": 2.6615200537894017, "gradient_norm_post_clip": 2.6615200537894017, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2238.0, "replay_loss": 0.47278396365647724, "replay_num_supervised_tokens": 45992, "total_loss": 0.5473826721312234, "replay_lambda": 1.0, "buffer_rows": 3584, "buffer_prompts": 264, "replay_draw_n_groups": 3520, "replay_draw_mean_ce": 0.1363531701705035, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 57, "epoch": 0, "batch_id": 3647, "num_supervised_tokens": 28839, "learning_rate": 1e-05, "batch_loss": 0.08175858542630034, "token_entropy": 0.10531814055616587, "gradient_norm_pre_clip": 4.314478927893975, "gradient_norm_post_clip": 4.314478927893975, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2750.0, "replay_loss": 0.41757436416337146, "replay_num_supervised_tokens": 45227, "total_loss": 0.4993329495896718, "replay_lambda": 1.0, "buffer_rows": 3648, "buffer_prompts": 264, "replay_draw_n_groups": 3584, "replay_draw_mean_ce": 0.04744747851757293, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 58, "epoch": 0, "batch_id": 3711, "num_supervised_tokens": 23554, "learning_rate": 1e-05, "batch_loss": 0.09600547357281491, "token_entropy": 0.11175808611391477, "gradient_norm_pre_clip": 3.1107049983550974, "gradient_norm_post_clip": 3.1107049983550974, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2607.0, "replay_loss": 0.4106678555472066, "replay_num_supervised_tokens": 46572, "total_loss": 0.5066733291200215, "replay_lambda": 1.0, "buffer_rows": 3712, "buffer_prompts": 267, "replay_draw_n_groups": 3648, "replay_draw_mean_ce": 0.43055364820692277, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 59, "epoch": 0, "batch_id": 3775, "num_supervised_tokens": 27043, "learning_rate": 1e-05, "batch_loss": 0.09036124338043001, "token_entropy": 0.0993040252912372, "gradient_norm_pre_clip": 2.3797346316811883, "gradient_norm_post_clip": 2.3797346316811883, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 766.0, "replay_loss": 0.31448013986749895, "replay_num_supervised_tokens": 42050, "total_loss": 0.40484138324792895, "replay_lambda": 1.0, "buffer_rows": 3776, "buffer_prompts": 268, "replay_draw_n_groups": 3712, "replay_draw_mean_ce": 0.1644673847016834, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 60, "epoch": 0, "batch_id": 3839, "num_supervised_tokens": 29023, "learning_rate": 1e-05, "batch_loss": 0.07081530877348485, "token_entropy": 0.07956561642010536, "gradient_norm_pre_clip": 1.8311243502895473, "gradient_norm_post_clip": 1.8311243502895473, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 836.0, "replay_loss": 0.3945397093781242, "replay_num_supervised_tokens": 43972, "total_loss": 0.46535501815160907, "replay_lambda": 1.0, "buffer_rows": 3840, "buffer_prompts": 268, "replay_draw_n_groups": 3776, "replay_draw_mean_ce": 0.05673548819004804, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.4368256579357, "parameter_l2_from_init": 9.398854844289806, "relative_parameter_l2_from_init": 0.008456755672970964, "update_norm": 1.0795772062493278, "relative_update_norm": 0.0009713347455535783}
+{"global_step": 61, "epoch": 0, "batch_id": 3903, "num_supervised_tokens": 29055, "learning_rate": 1e-05, "batch_loss": 0.08734134451949828, "token_entropy": 0.08346095501176463, "gradient_norm_pre_clip": 2.3529774061806115, "gradient_norm_post_clip": 2.3529774061806115, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1139.0, "replay_loss": 0.3928482148899561, "replay_num_supervised_tokens": 44491, "total_loss": 0.48018955940945435, "replay_lambda": 1.0, "buffer_rows": 3904, "buffer_prompts": 269, "replay_draw_n_groups": 3840, "replay_draw_mean_ce": 0.2942938333029275, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 62, "epoch": 0, "batch_id": 3967, "num_supervised_tokens": 27553, "learning_rate": 1e-05, "batch_loss": 0.08743599059147895, "token_entropy": 0.08063614028804952, "gradient_norm_pre_clip": 2.3354252228007284, "gradient_norm_post_clip": 2.3354252228007284, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 810.0, "replay_loss": 0.3736226338135205, "replay_num_supervised_tokens": 44113, "total_loss": 0.46105862440499945, "replay_lambda": 1.0, "buffer_rows": 3968, "buffer_prompts": 270, "replay_draw_n_groups": 3904, "replay_draw_mean_ce": 0.5994729841411307, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 63, "epoch": 0, "batch_id": 4031, "num_supervised_tokens": 29373, "learning_rate": 1e-05, "batch_loss": 0.08545683009324033, "token_entropy": 0.0779901017890426, "gradient_norm_pre_clip": 4.2919554985796875, "gradient_norm_post_clip": 4.2919554985796875, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.31159049562129615, "replay_num_supervised_tokens": 44741, "total_loss": 0.3970473257145365, "replay_lambda": 1.0, "buffer_rows": 4032, "buffer_prompts": 270, "replay_draw_n_groups": 3968, "replay_draw_mean_ce": 0.15228283869755732, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 64, "epoch": 0, "batch_id": 4095, "num_supervised_tokens": 30287, "learning_rate": 1e-05, "batch_loss": 0.09689587224049784, "token_entropy": 0.09718741111718845, "gradient_norm_pre_clip": 2.6949889122567865, "gradient_norm_post_clip": 2.6949889122567865, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.3272590228654568, "replay_num_supervised_tokens": 43988, "total_loss": 0.4241548951059546, "replay_lambda": 1.0, "buffer_rows": 4096, "buffer_prompts": 270, "replay_draw_n_groups": 4032, "replay_draw_mean_ce": 0.07971867271091627, "replay_draw_frac_unobserved": 0.0}
+{"global_step": 65, "epoch": 0, "batch_id": 4105, "num_supervised_tokens": 4748, "learning_rate": 1e-05, "batch_loss": 0.2045841271094209, "token_entropy": 0.22731851547228357, "gradient_norm_pre_clip": 3.1958495039384514, "gradient_norm_post_clip": 3.1958495039384514, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 846.0, "replay_loss": 0.3133787680156018, "replay_num_supervised_tokens": 6416, "total_loss": 0.5179628951250227, "replay_lambda": 1.0, "buffer_rows": 4106, "buffer_prompts": 270, "replay_draw_n_groups": 4096, "replay_draw_mean_ce": 0.2715442112513951, "replay_draw_frac_unobserved": 0.0, "parameter_l2_norm": 1111.4362000804929, "parameter_l2_from_init": 9.563821874373094, "relative_parameter_l2_from_init": 0.008605187145806995, "update_norm": 0.4898771379922531, "relative_update_norm": 0.0004407604664638195}

training/llama3b_sft_rp_ce_lam1/provenance.json

@@ -0,0 +1,25 @@
+{
+  "base_model": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "dataset": "llama3b_code_sft_data",
+  "dataset_sha256": "69ec8901893709bcba86b78d434c5995f523c783916edffa0129aed8649040de",
+  "durable_jsonl_sha256": "757270437abe9b280a7104cb5b13223f32d782444e42bda77e5907c4c1b5d7e5",
+  "effective_batch_size": 256,
+  "max_seq_length": 4096,
+  "metrics_log_every": 10,
+  "micro_batch_size": 1,
+  "model_name": "llama3b",
+  "num_checkpoints": 10,
+  "order": "ordered",
+  "replay_lambda": 1.0,
+  "replay_strategy": "ce",
+  "resume_generation": 1,
+  "resume_helper_sha256": "e30bddb7e051f29d4a16700b2b5dcec4e572785be673ab7d42480c0c4f3f0707",
+  "resume_step": 0,
+  "run_sft_sha256": "f4ee1ff82e7da742db079178b8d3695d2142f1db42096b0f1dcb5d837baea263",
+  "run_sweep_sha256": "aec506941a4ca39773d465ce72586294a32e9b9d1761902201976fa8d28386fc",
+  "save_optimizer_every_n_checkpoints": 5,
+  "save_optimizer_first_checkpoint": true,
+  "schema_version": 1,
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "world_size": 4
+}

training/llama3b_sft_rp_ce_lam1/sft_config.json

@@ -0,0 +1,56 @@
+{
+  "generation_logs_dir": null,
+  "dataset": "llama3b_code_sft_data",
+  "positives_file": "/mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl",
+  "model_path": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "output_dir": "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922",
+  "lr": 1e-05,
+  "batch_size": 1,
+  "effective_batch_size": 256,
+  "weight_decay": 0.01,
+  "warmup_ratio": 0.0,
+  "schedule": "constant",
+  "max_seq_length": 4096,
+  "seed": 42,
+  "bf16": true,
+  "attn_impl": "auto",
+  "precision": "mixed_bf16",
+  "grad_clip": 0.0,
+  "loss_agg_mode": "token-mean",
+  "ordered": true,
+  "within_step_shuffle": false,
+  "system_prompt_mode": "from_record",
+  "score_threshold": 0.0,
+  "drop_all_pass_groups": false,
+  "max_rl_step": null,
+  "num_epochs": 1,
+  "replay_strategy": "ce",
+  "replay_lambda": 1.0,
+  "replay_batch_size": 1,
+  "replay_cooldown_step": 3.0,
+  "replay_ce_tau": 1.0,
+  "replay_ce_init": "insert",
+  "replay_group_size": 8,
+  "replay_seed": null,
+  "max_train_steps": null,
+  "num_checkpoints": 10,
+  "checkpoint_after_resume_steps": null,
+  "save_optimizer_state": true,
+  "prune_old_optimizer": true,
+  "save_optimizer_every_n_ckpts": 5,
+  "save_optimizer_first_ckpt": true,
+  "snapshot_dtype": "float32",
+  "snapshot_device": "cpu",
+  "metrics_log_every": 10,
+  "ddp_timeout_seconds": 3600,
+  "memory_canary_longest_first": false,
+  "log_cuda_memory": true,
+  "save_final_model": true,
+  "wandb_project": "rl_forgetting_sft",
+  "wandb_entity": "jingyanshen-new-york-university",
+  "wandb_run_name": "llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922_segment_g1",
+  "wandb_group": "code_sft_20260922_llama3b",
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "resume_from": null,
+  "grad_accum_steps": 64
+}

training/llama3b_sft_rp_ce_lam1/training_complete.json

@@ -0,0 +1,7 @@
+{
+  "azure_run_id": "busy_drop_hjxskmhkmq",
+  "completed_at": "2026-09-23T01:08:45.443877+00:00",
+  "host": "node-0",
+  "schema_version": 1,
+  "status": "training_complete"
+}

training/llama3b_sft_rp_ce_lam1/wandb/run-20260923_003243-ihjigugm/files/config.yaml

@@ -0,0 +1,253 @@
+_wandb:
+    value:
+        cli_version: 0.28.0
+        e:
+            ctw3dwp25rgky3k4s74780wrez01ki0h:
+                args:
+                    - --dataset
+                    - llama3b_code_sft_data
+                    - --replay_strategy
+                    - ce
+                    - --replay_lambda
+                    - "1.0"
+                    - --replay_cooldown_step
+                    - "3"
+                    - --replay_ce_tau
+                    - "1.0"
+                    - --replay_ce_init
+                    - insert
+                    - --replay_group_size
+                    - "8"
+                    - --model_path
+                    - /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+                    - --output_dir
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922
+                    - --ordered
+                    - --system_prompt_mode
+                    - from_record
+                    - --lr
+                    - "1e-5"
+                    - --batch_size
+                    - "1"
+                    - --effective_batch_size
+                    - "256"
+                    - --max_seq_length
+                    - "4096"
+                    - --ddp_timeout_seconds
+                    - "3600"
+                    - --schedule
+                    - constant
+                    - --warmup_ratio
+                    - "0"
+                    - --num_epochs
+                    - "1"
+                    - --seed
+                    - "42"
+                    - --log_cuda_memory
+                    - --num_checkpoints
+                    - "10"
+                    - --save_optimizer_every_n_ckpts
+                    - "5"
+                    - --prune_old_optimizer
+                    - --save_optimizer_first_ckpt
+                    - --metrics_log_every
+                    - "10"
+                    - --precision
+                    - mixed_bf16
+                    - --grad_clip
+                    - "0"
+                    - --snapshot_device
+                    - cpu
+                    - --snapshot_dtype
+                    - float32
+                    - --wandb_project
+                    - rl_forgetting_sft
+                    - --wandb_entity
+                    - jingyanshen-new-york-university
+                    - --wandb_group
+                    - code_sft_20260922_llama3b
+                    - --wandb_run_name
+                    - llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922_segment_g1
+                    - --wandb_run_id_file
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+                codePath: sft/run_sft.py
+                codePathLocal: sft/run_sft.py
+                cpu_count: 96
+                cpu_count_logical: 96
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "30601604947968"
+                        used: "1136683208704"
+                email: js15262@nyu.edu
+                executable: /tmp/rlf-code-runtime/bin/python
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-9c1b9503-5afd-ab8c-0e01-6c8337da0a53
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-e3b2a910-e616-ad55-7fd8-4048d0980905
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-b401c410-715e-d3f8-1b7d-96f4e1500779
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-4463b329-8c56-f918-cf39-5cff3953dd8d
+                host: node-0
+                memory:
+                    total: "2008079081472"
+                os: Linux-6.8.0-1062-azure-x86_64-with-glibc2.35
+                program: /scratch/amlt_code/sft/run_sft.py
+                python: CPython 3.10.21
+                root: /mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922
+                startedAt: "2026-09-23T00:32:43.665824Z"
+                writerId: ctw3dwp25rgky3k4s74780wrez01ki0h
+        m: []
+        python_version: 3.10.21
+        t:
+            "1":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "2":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "3":
+                - 2
+                - 13
+                - 14
+                - 16
+                - 61
+            "4": 3.10.21
+            "5": 0.28.0
+            "6": 4.57.6
+            "12": 0.28.0
+            "13": linux-x86_64
+attn_impl:
+    value: auto
+batch_size:
+    value: 1
+bf16:
+    value: true
+checkpoint_after_resume_steps:
+    value: null
+dataset:
+    value: llama3b_code_sft_data
+ddp_timeout_seconds:
+    value: 3600
+drop_all_pass_groups:
+    value: false
+effective_batch_size:
+    value: 256
+generation_logs_dir:
+    value: null
+grad_accum_steps:
+    value: 64
+grad_clip:
+    value: 0
+log_cuda_memory:
+    value: true
+loss_agg_mode:
+    value: token-mean
+lr:
+    value: 1e-05
+max_rl_step:
+    value: null
+max_seq_length:
+    value: 4096
+max_train_steps:
+    value: null
+memory_canary_longest_first:
+    value: false
+metrics_log_every:
+    value: 10
+model_path:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+num_checkpoints:
+    value: 10
+num_epochs:
+    value: 1
+ordered:
+    value: true
+output_dir:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922
+positives_file:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl
+precision:
+    value: mixed_bf16
+prune_old_optimizer:
+    value: true
+replay_batch_size:
+    value: 1
+replay_ce_init:
+    value: insert
+replay_ce_tau:
+    value: 1
+replay_cooldown_step:
+    value: 3
+replay_group_size:
+    value: 8
+replay_lambda:
+    value: 1
+replay_seed:
+    value: null
+replay_strategy:
+    value: ce
+resume_from:
+    value: null
+save_final_model:
+    value: true
+save_optimizer_every_n_ckpts:
+    value: 5
+save_optimizer_first_ckpt:
+    value: true
+save_optimizer_state:
+    value: true
+schedule:
+    value: constant
+score_threshold:
+    value: 0
+seed:
+    value: 42
+snapshot_device:
+    value: cpu
+snapshot_dtype:
+    value: float32
+system_prompt_mode:
+    value: from_record
+wandb_entity:
+    value: jingyanshen-new-york-university
+wandb_group:
+    value: code_sft_20260922_llama3b
+wandb_project:
+    value: rl_forgetting_sft
+wandb_run_id_file:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+wandb_run_name:
+    value: llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922_segment_g1
+warmup_ratio:
+    value: 0
+weight_decay:
+    value: 0.01
+within_step_shuffle:
+    value: false

training/llama3b_sft_rp_ce_lam1/wandb/run-20260923_003243-ihjigugm/files/requirements.txt

@@ -0,0 +1,266 @@
+pyzmq==27.2.0
+MarkupSafe==3.0.3
+grpcio==1.83.1
+mistral_common==1.11.7
+msgpack==1.2.2
+click==8.5.0
+opentelemetry-exporter-otlp==1.26.0
+gitdb==4.0.12
+pyfiglet==1.0.2
+sqlparse==0.6.0
+tensordict==0.6.2
+httpcore==1.0.9
+gguf==0.19.0
+nvidia-cuda-cupti-cu12==12.4.127
+virtualenv==21.7.9
+SQLAlchemy==2.0.52
+docker==7.2.0
+google-auth==2.58.0
+transformers==4.57.6
+cloudpickle==3.1.2
+opentelemetry-sdk==1.26.0
+detect-installer==0.2.1
+opentelemetry-semantic-conventions==0.47b0
+agent-detector==2.0.0
+py-cpuinfo==9.0.0
+accelerate==1.15.0
+airportsdata==20260315
+pybind11==3.1.0
+xxhash==4.0.1
+cffi==2.1.1
+tzdata==2026.3
+diskcache==5.6.3
+vllm==0.8.5
+mlflow-skinny==3.16.0
+distlib==0.4.3
+msgspec==0.21.1
+Mako==1.4.1
+depyf==0.18.0
+truststore==0.10.4
+tabulate==0.9.0
+smmap==5.0.3
+pyparsing==3.3.2
+opentelemetry-exporter-prometheus==0.47b0
+gunicorn==26.2.0
+fastar==0.12.0
+scipy==1.15.3
+llguidance==0.7.30
+outlines==0.1.11
+watchfiles==1.2.0
+httpx==0.28.1
+prettytable==3.18.0
+torchaudio==2.6.0
+tomli==2.4.1
+math-verify==0.9.0
+pyasn1_modules==0.4.2
+skops==0.14.0
+compressed-tensors==0.9.3
+einops==0.8.2
+graphene==3.4.3
+flask-cors==6.0.5
+cuda-pathfinder==1.8.1
+scikit-learn==1.7.2
+reasoning_gym==0.1.25
+astor==0.8.1
+openai==3.13.0
+alembic==1.19.2
+partial-json-parser==0.2.1.1.post7
+dill==0.4.1
+greenlet==3.5.5
+pydantic-extra-types==2.11.1
+Werkzeug==3.1.8
+email-validator==2.3.0
+pycountry==26.2.16
+pip==26.2.1
+arckit==0.1.0
+Deprecated==1.3.1
+nvidia-cusparse-cu12==12.3.1.170
+bfi==1.0.4
+packaging==26.3
+databricks-sdk==0.138.0
+flash_attn==2.7.4.post1
+psutil==7.2.2
+exceptiongroup==1.3.1
+smart_open==8.0.1
+idna==3.19
+cycler==0.12.1
+mlflow==3.16.0
+huggingface_hub==0.36.2
+pyarrow==25.0.1
+nvidia-nvjitlink-cu12==12.4.127
+protobuf==4.25.9
+zipp==4.1.0
+numpy==2.2.6
+pycosat==0.6.6
+mlflow-tracing==3.16.0
+filelock==3.32.6
+fastapi-cloud-cli==0.26.0
+annotated-types==0.8.0
+python-dotenv==1.2.3
+six==1.17.0
+outlines_core==0.1.26
+pillow==12.3.0
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cublas-cu12==12.4.5.8
+typing-inspection==0.4.4
+charset-normalizer==3.5.1
+frozenlist==1.8.0
+Jinja2==3.1.6
+latex2sympy2_extended==1.11.0
+cryptography==50.0.1
+py-spy==0.4.2
+networkx==3.4.2
+wrapt==2.4.1
+safetensors==0.8.0
+omegaconf==2.3.1
+websockets==16.1.1
+shellingham==1.5.4
+itsdangerous==2.2.0
+torchvision==0.21.0
+pandas==2.3.3
+pydantic-settings==2.15.0
+aiohttp-cors==0.8.1
+nvidia-cusparselt-cu12==0.6.2
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+attrs==26.1.0
+python-discovery==1.6.0
+cellpylib==2.4.0
+annotated-doc==0.0.5
+graphql-core==3.2.12
+typing_extensions==4.16.0
+sniffio==1.3.1
+python-dateutil==2.9.0.post0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+interegular==0.3.3
+nvidia-cuda-nvrtc-cu12==12.4.127
+python-json-logger==4.2.0
+nvidia-cufft-cu12==11.2.1.3
+wandb==0.28.0
+mdurl==0.1.2
+anyio==4.15.1
+wcwidth==0.8.3
+opentelemetry-api==1.26.0
+peft==0.20.0
+uvicorn==0.52.4
+pydantic_core==2.46.5
+opencv-python-headless==5.0.0.93
+proto-plus==1.28.2
+llvmlite==0.44.0
+numba==0.61.2
+propcache==0.5.2
+starlette==1.6.0
+opentelemetry-proto==1.26.0
+sympy==1.13.1
+tqdm==4.70.1
+drawsvg==2.4.2
+blake3==1.0.9
+urllib3==2.7.0
+pytz==2026.3.post1
+opencensus-context==0.1.3
+blinker==1.9.0
+fsspec==2026.6.0
+opentelemetry-exporter-otlp-proto-http==1.26.0
+hf-xet==1.6.0
+jsonschema-specifications==2025.9.1
+zss==1.2.0
+magiccube==0.3.0
+matplotlib==3.10.9
+referencing==0.37.0
+async-timeout==5.0.1
+fastapi-cli==0.0.32
+contourpy==1.3.2
+platformdirs==4.11.8
+huey==3.4.0
+aiohttp==3.14.3
+certifi==2026.7.22
+fonttools==4.65.0
+GitPython==3.1.62
+rich-toolkit==0.20.5
+multiprocess==0.70.19
+sentencepiece==0.2.2
+verl==0.4.1.dev0
+codetiming==1.4.0
+Pygments==2.21.0
+python-multipart==0.0.32
+joblib==1.6.0
+pyasn1==0.6.4
+markdown-it-py==4.2.0
+jiter==0.16.0
+aiosignal==1.4.0
+dnspython==2.8.0
+httpx2==2.12.0
+kiwisolver==1.5.1
+lm-format-enforcer==0.10.12
+setuptools==80.9.0
+jsonschema==4.26.0
+Flask==3.1.3
+orjson==3.12.0
+pycparser==3.0
+lark==1.2.2
+cupy-cuda12x==14.2.0
+importlib_metadata==8.0.0
+nest-asyncio==1.6.0
+PyYAML==6.0.3
+nvidia-cudnn-cu12==9.1.0.70
+wheel==0.47.0
+uvloop==0.22.1
+torch==2.6.0
+triton==3.2.0
+colorful==0.5.8
+pylatexenc==2.11
+typer==0.27.2
+google-api-core==2.30.3
+ninja==1.13.2
+mpmath==1.3.0
+graphql-relay==3.2.0
+hydra-core==1.3.6
+rich==15.0.0
+antlr4-python3-runtime==4.9.3
+cachetools==7.1.8
+prometheus-fastapi-instrumentator==8.1.0
+tokenizers==0.22.2
+rignore==0.8.1
+regex==2026.9.10
+datasets==5.0.1
+fastapi==0.141.1
+nvidia-curand-cu12==10.3.5.147
+requests==2.34.2
+tiktoken==0.14.0
+nvidia-nvtx-cu12==12.4.127
+nvidia-cusolver-cu12==11.6.1.9
+sentry-sdk==2.69.1
+multidict==6.8.0
+opentelemetry-semantic-conventions-ai==0.4.13
+aiohappyeyeballs==2.7.1
+rpds-py==0.30.0
+httpcore2==2.12.0
+prometheus_client==0.26.0
+torchdata==0.11.0
+xformers==0.0.29.post2
+ray==2.47.1
+conda-pack==0.9.2
+opencensus==0.11.4
+googleapis-common-protos==1.75.0
+yarl==1.24.5
+h11==0.16.0
+nvidia-nccl-cu12==2.21.5
+pydantic==2.13.5
+threadpoolctl==3.6.0
+httptools==0.8.0
+xgrammar==0.1.18
+platformdirs==4.2.2
+more-itertools==10.3.0
+inflect==7.3.1
+jaraco.collections==5.1.0
+typing_extensions==4.12.2
+jaraco.text==3.12.1
+zipp==3.19.2
+tomli==2.0.1
+jaraco.functools==4.0.1
+typeguard==4.3.0
+packaging==24.2
+autocommand==2.2.2
+wheel==0.45.1
+importlib_metadata==8.0.0
+jaraco.context==5.3.0
+backports.tarfile==1.2.0

training/llama3b_sft_rp_ce_lam1/wandb/run-20260923_003243-ihjigugm/files/wandb-metadata.json

@@ -0,0 +1,125 @@
+{
+  "os": "Linux-6.8.0-1062-azure-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.21",
+  "startedAt": "2026-09-23T00:32:43.665824Z",
+  "args": [
+    "--dataset",
+    "llama3b_code_sft_data",
+    "--replay_strategy",
+    "ce",
+    "--replay_lambda",
+    "1.0",
+    "--replay_cooldown_step",
+    "3",
+    "--replay_ce_tau",
+    "1.0",
+    "--replay_ce_init",
+    "insert",
+    "--replay_group_size",
+    "8",
+    "--model_path",
+    "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+    "--output_dir",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922",
+    "--ordered",
+    "--system_prompt_mode",
+    "from_record",
+    "--lr",
+    "1e-5",
+    "--batch_size",
+    "1",
+    "--effective_batch_size",
+    "256",
+    "--max_seq_length",
+    "4096",
+    "--ddp_timeout_seconds",
+    "3600",
+    "--schedule",
+    "constant",
+    "--warmup_ratio",
+    "0",
+    "--num_epochs",
+    "1",
+    "--seed",
+    "42",
+    "--log_cuda_memory",
+    "--num_checkpoints",
+    "10",
+    "--save_optimizer_every_n_ckpts",
+    "5",
+    "--prune_old_optimizer",
+    "--save_optimizer_first_ckpt",
+    "--metrics_log_every",
+    "10",
+    "--precision",
+    "mixed_bf16",
+    "--grad_clip",
+    "0",
+    "--snapshot_device",
+    "cpu",
+    "--snapshot_dtype",
+    "float32",
+    "--wandb_project",
+    "rl_forgetting_sft",
+    "--wandb_entity",
+    "jingyanshen-new-york-university",
+    "--wandb_group",
+    "code_sft_20260922_llama3b",
+    "--wandb_run_name",
+    "llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922_segment_g1",
+    "--wandb_run_id_file",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt"
+  ],
+  "program": "/scratch/amlt_code/sft/run_sft.py",
+  "codePath": "sft/run_sft.py",
+  "codePathLocal": "sft/run_sft.py",
+  "email": "js15262@nyu.edu",
+  "root": "/mnt/output/projects/rl-forgetting/amlt-results/7209851795.96752-c0f24590-b615/llama3b_code_ordered_lr1e-5_bs256_seed42_rpce_lam1.0_gcr_20260922",
+  "host": "node-0",
+  "executable": "/tmp/rlf-code-runtime/bin/python",
+  "cpu_count": 96,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "30601604947968",
+      "used": "1136683208704"
+    }
+  },
+  "memory": {
+    "total": "2008079081472"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-9c1b9503-5afd-ab8c-0e01-6c8337da0a53"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-e3b2a910-e616-ad55-7fd8-4048d0980905"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-b401c410-715e-d3f8-1b7d-96f4e1500779"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-4463b329-8c56-f918-cf39-5cff3953dd8d"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "ctw3dwp25rgky3k4s74780wrez01ki0h"
+}

training/llama3b_sft_rp_ce_lam1/wandb/run-20260923_003243-ihjigugm/files/wandb-summary.json

@@ -0,0 +1 @@
+{"cuda_peak_reserved_gib":73.154296875,"replay_lambda":1,"_timestamp":1.7901255482349696e+09,"cuda_peak_allocated_gib":71.87305641174316,"epoch":0,"replay_loss":0.3133787680156018,"token_entropy":0.22731851547228357,"replay_draw_frac_unobserved":0,"batch_loss":0.2045841271094209,"_step":65,"total_loss":0.5179628951250227,"learning_rate":1e-05,"buffer_prompts":270,"replay_num_supervised_tokens":6416,"_runtime":2135.455804606,"parameter_l2_norm":1111.4362000804929,"relative_update_norm":0.0004407604664638195,"update_norm":0.4898771379922531,"batch_id":4105,"_wandb":{"runtime":2135},"buffer_rows":4106,"cuda_current_reserved_gib":73.154296875,"num_supervised_tokens":4748,"replay_draw_n_groups":4096,"relative_parameter_l2_from_init":0.008605187145806995,"cuda_current_allocated_gib":47.93620491027832,"replay_draw_mean_ce":0.2715442112513951,"gradient_norm_post_clip":3.1958495039384514,"max_sequence_length":846,"global_step":65,"parameter_l2_from_init":9.563821874373094,"gradient_norm_pre_clip":3.1958495039384514}

training/llama3b_sft_rp_ce_lam1/wandb_run_id_generation_1.txt

@@ -0,0 +1 @@
+ihjigugm

training/llama3b_sft_rp_ce_lam1/wandb_segments/generation_1.json

@@ -0,0 +1,8 @@
+{
+  "attempt_id": "busy_drop_hjxskmhkmq|node-0|05c1ba9fa3af4efd9717154ba156013d",
+  "created_at": "2026-09-23T00:30:20.843204+00:00",
+  "generation": 1,
+  "resume_step": 0,
+  "schema_version": 1,
+  "wandb_run_id": "ihjigugm"
+}

training/llama3b_sft_rp_hard_lam1/metrics/train_metrics.jsonl

@@ -0,0 +1,65 @@
+{"global_step": 1, "epoch": 0, "batch_id": 63, "num_supervised_tokens": 97674, "learning_rate": 1e-05, "batch_loss": 1.3036121778480152, "token_entropy": 1.3033737616757028, "gradient_norm_pre_clip": 1.6957898516134071, "gradient_norm_post_clip": 1.6957898516134071, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.1171875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.1171875, "max_sequence_length": 3030.0, "replay_loss": NaN, "replay_num_supervised_tokens": 0, "total_loss": 1.3036121778480152, "replay_lambda": 1.0, "buffer_rows": 64, "buffer_prompts": 54}
+{"global_step": 2, "epoch": 0, "batch_id": 127, "num_supervised_tokens": 92723, "learning_rate": 1e-05, "batch_loss": 0.6894052104700573, "token_entropy": 0.6533522015885673, "gradient_norm_pre_clip": 3.403035121780755, "gradient_norm_post_clip": 3.403035121780755, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 3030.0, "replay_loss": 1.5842116603789869, "replay_num_supervised_tokens": 101633, "total_loss": 2.273616870849044, "replay_lambda": 1.0, "buffer_rows": 128, "buffer_prompts": 107, "replay_draw_n_groups": 54, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00819672131147541, "replay_draw_max_q": 0.028688524590163935}
+{"global_step": 3, "epoch": 0, "batch_id": 191, "num_supervised_tokens": 92641, "learning_rate": 1e-05, "batch_loss": 0.7249780930925739, "token_entropy": 0.7564943337296113, "gradient_norm_pre_clip": 8.240382772017785, "gradient_norm_post_clip": 8.240382772017785, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 2607.0, "replay_loss": 1.138569984507642, "replay_num_supervised_tokens": 97073, "total_loss": 1.8635480776002158, "replay_lambda": 1.0, "buffer_rows": 192, "buffer_prompts": 127, "replay_draw_n_groups": 107, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0023148148148148147, "replay_draw_max_q": 0.016203703703703703}
+{"global_step": 4, "epoch": 0, "batch_id": 255, "num_supervised_tokens": 87718, "learning_rate": 1e-05, "batch_loss": 0.9069567020562825, "token_entropy": 0.9236716085600215, "gradient_norm_pre_clip": 2.311716663371294, "gradient_norm_post_clip": 2.311716663371294, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 3030.0, "replay_loss": 0.9383650432734875, "replay_num_supervised_tokens": 96674, "total_loss": 1.8453217453297701, "replay_lambda": 1.0, "buffer_rows": 256, "buffer_prompts": 148, "replay_draw_n_groups": 148, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0017761989342806395, "replay_draw_max_q": 0.012433392539964476}
+{"global_step": 5, "epoch": 0, "batch_id": 319, "num_supervised_tokens": 82516, "learning_rate": 1e-05, "batch_loss": 0.8517448556390899, "token_entropy": 0.7894902048088309, "gradient_norm_pre_clip": 2.918674719660005, "gradient_norm_post_clip": 2.918674719660005, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 2066.0, "replay_loss": 1.1538019654523832, "replay_num_supervised_tokens": 96500, "total_loss": 2.005546821091473, "replay_lambda": 1.0, "buffer_rows": 320, "buffer_prompts": 168, "replay_draw_n_groups": 195, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0013605442176870747, "replay_draw_max_q": 0.009523809523809525}
+{"global_step": 6, "epoch": 0, "batch_id": 383, "num_supervised_tokens": 78503, "learning_rate": 1e-05, "batch_loss": 1.1523821343679348, "token_entropy": 1.1095556474883952, "gradient_norm_pre_clip": 2.1521868853482555, "gradient_norm_post_clip": 2.1521868853482555, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 2750.0, "replay_loss": 0.7186098322880632, "replay_num_supervised_tokens": 90632, "total_loss": 1.870991966655998, "replay_lambda": 1.0, "buffer_rows": 384, "buffer_prompts": 182, "replay_draw_n_groups": 244, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0011376564277588168, "replay_draw_max_q": 0.007963594994311717}
+{"global_step": 7, "epoch": 0, "batch_id": 447, "num_supervised_tokens": 75606, "learning_rate": 1e-05, "batch_loss": 1.1958866439121358, "token_entropy": 1.1916121355021494, "gradient_norm_pre_clip": 1.3933443554984788, "gradient_norm_post_clip": 1.3933443554984788, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.724609375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.724609375, "max_sequence_length": 2916.0, "replay_loss": 0.9360977427933538, "replay_num_supervised_tokens": 95720, "total_loss": 2.1319843867054895, "replay_lambda": 1.0, "buffer_rows": 448, "buffer_prompts": 190, "replay_draw_n_groups": 288, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0009727626459143969, "replay_draw_max_q": 0.006809338521400778}
+{"global_step": 8, "epoch": 0, "batch_id": 511, "num_supervised_tokens": 65460, "learning_rate": 1e-05, "batch_loss": 0.6353756713889496, "token_entropy": 0.6481193460312379, "gradient_norm_pre_clip": 1.7344331328151128, "gradient_norm_post_clip": 1.7344331328151128, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 1.2716953413732888, "replay_num_supervised_tokens": 98084, "total_loss": 1.9070710127622383, "replay_lambda": 1.0, "buffer_rows": 512, "buffer_prompts": 197, "replay_draw_n_groups": 337, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0008396305625524769, "replay_draw_max_q": 0.005877413937867338}
+{"global_step": 9, "epoch": 0, "batch_id": 575, "num_supervised_tokens": 60380, "learning_rate": 1e-05, "batch_loss": 0.6590584813489765, "token_entropy": 0.6792807680125582, "gradient_norm_pre_clip": 1.8233986386032899, "gradient_norm_post_clip": 1.8233986386032899, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2750.0, "replay_loss": 1.2841337347297115, "replay_num_supervised_tokens": 94772, "total_loss": 1.943192216078688, "replay_lambda": 1.0, "buffer_rows": 576, "buffer_prompts": 200, "replay_draw_n_groups": 386, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0007473841554559044, "replay_draw_max_q": 0.00523168908819133}
+{"global_step": 10, "epoch": 0, "batch_id": 639, "num_supervised_tokens": 60964, "learning_rate": 1e-05, "batch_loss": 0.5182409229799622, "token_entropy": 0.5297220471446403, "gradient_norm_pre_clip": 1.4874648305982905, "gradient_norm_post_clip": 1.4874648305982905, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1236.0, "replay_loss": 0.6941899064680067, "replay_num_supervised_tokens": 82197, "total_loss": 1.2124308294479689, "replay_lambda": 1.0, "buffer_rows": 640, "buffer_prompts": 204, "replay_draw_n_groups": 429, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0006802721088435374, "replay_draw_max_q": 0.004761904761904762, "parameter_l2_norm": 1111.3996831538223, "parameter_l2_from_init": 2.2674778109384905, "relative_parameter_l2_from_init": 0.0020401959769215687, "update_norm": 2.2674778109384905, "relative_update_norm": 0.002040200159589808}
+{"global_step": 11, "epoch": 0, "batch_id": 703, "num_supervised_tokens": 54290, "learning_rate": 1e-05, "batch_loss": 0.4804143711998661, "token_entropy": 0.4635157912404174, "gradient_norm_pre_clip": 1.555783719765749, "gradient_norm_post_clip": 1.555783719765749, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.9232692695129833, "replay_num_supervised_tokens": 82901, "total_loss": 1.4036836407128495, "replay_lambda": 1.0, "buffer_rows": 704, "buffer_prompts": 206, "replay_draw_n_groups": 469, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0006357279084551812, "replay_draw_max_q": 0.004450095359186269}
+{"global_step": 12, "epoch": 0, "batch_id": 767, "num_supervised_tokens": 54744, "learning_rate": 1e-05, "batch_loss": 0.4308902089504859, "token_entropy": 0.4122668993014891, "gradient_norm_pre_clip": 1.6758358001070104, "gradient_norm_post_clip": 1.6758358001070104, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.830378300718032, "replay_num_supervised_tokens": 80556, "total_loss": 1.261268509668518, "replay_lambda": 1.0, "buffer_rows": 768, "buffer_prompts": 209, "replay_draw_n_groups": 516, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.0005889281507656066, "replay_draw_max_q": 0.004122497055359246}
+{"global_step": 13, "epoch": 0, "batch_id": 831, "num_supervised_tokens": 48778, "learning_rate": 1e-05, "batch_loss": 0.43635445097923775, "token_entropy": 0.4280812724782628, "gradient_norm_pre_clip": 1.4789077974128464, "gradient_norm_post_clip": 1.4789077974128464, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.870544365965115, "replay_num_supervised_tokens": 83574, "total_loss": 1.3068988169443527, "replay_lambda": 1.0, "buffer_rows": 832, "buffer_prompts": 212, "replay_draw_n_groups": 555, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.0005543237250554324, "replay_draw_max_q": 0.003880266075388027}
+{"global_step": 14, "epoch": 0, "batch_id": 895, "num_supervised_tokens": 48298, "learning_rate": 1e-05, "batch_loss": 0.40962091418401764, "token_entropy": 0.41424787274068064, "gradient_norm_pre_clip": 1.498156110629564, "gradient_norm_post_clip": 1.498156110629564, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.8145722119358187, "replay_num_supervised_tokens": 78258, "total_loss": 1.2241931261198364, "replay_lambda": 1.0, "buffer_rows": 896, "buffer_prompts": 214, "replay_draw_n_groups": 598, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.0005197505197505198, "replay_draw_max_q": 0.0036382536382536385}
+{"global_step": 15, "epoch": 0, "batch_id": 959, "num_supervised_tokens": 44942, "learning_rate": 1e-05, "batch_loss": 0.35590592526278386, "token_entropy": 0.375494248939926, "gradient_norm_pre_clip": 1.5528790251083038, "gradient_norm_post_clip": 1.5528790251083038, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.6479422510162146, "replay_num_supervised_tokens": 80854, "total_loss": 1.0038481762789986, "replay_lambda": 1.0, "buffer_rows": 960, "buffer_prompts": 216, "replay_draw_n_groups": 648, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.0004880429477794046, "replay_draw_max_q": 0.003416300634455832}
+{"global_step": 16, "epoch": 0, "batch_id": 1023, "num_supervised_tokens": 40349, "learning_rate": 1e-05, "batch_loss": 0.34156892515286663, "token_entropy": 0.36106854978406966, "gradient_norm_pre_clip": 1.4708522113462283, "gradient_norm_post_clip": 1.4708522113462283, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2916.0, "replay_loss": 1.005199035740484, "replay_num_supervised_tokens": 75965, "total_loss": 1.3467679608933507, "replay_lambda": 1.0, "buffer_rows": 1024, "buffer_prompts": 218, "replay_draw_n_groups": 689, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.0004621072088724584, "replay_draw_max_q": 0.003234750462107209}
+{"global_step": 17, "epoch": 0, "batch_id": 1087, "num_supervised_tokens": 41901, "learning_rate": 1e-05, "batch_loss": 0.32768561296423354, "token_entropy": 0.3415488503047047, "gradient_norm_pre_clip": 1.4525548212619506, "gradient_norm_post_clip": 1.4525548212619506, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.9587329648770275, "replay_num_supervised_tokens": 80828, "total_loss": 1.2864185778412611, "replay_lambda": 1.0, "buffer_rows": 1088, "buffer_prompts": 223, "replay_draw_n_groups": 731, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00043975373790677223, "replay_draw_max_q": 0.0030782761653474055}
+{"global_step": 18, "epoch": 0, "batch_id": 1151, "num_supervised_tokens": 37655, "learning_rate": 1e-05, "batch_loss": 0.35113206156030496, "token_entropy": 0.35504907185837276, "gradient_norm_pre_clip": 1.5091842898273706, "gradient_norm_post_clip": 1.5091842898273706, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.8456252570668018, "replay_num_supervised_tokens": 70395, "total_loss": 1.1967573186271068, "replay_lambda": 1.0, "buffer_rows": 1152, "buffer_prompts": 225, "replay_draw_n_groups": 775, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.00041631973355537054, "replay_draw_max_q": 0.0029142381348875937}
+{"global_step": 19, "epoch": 0, "batch_id": 1215, "num_supervised_tokens": 38084, "learning_rate": 1e-05, "batch_loss": 0.29364214168086594, "token_entropy": 0.2940933506299237, "gradient_norm_pre_clip": 1.5378169310511058, "gradient_norm_post_clip": 1.5378169310511058, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 1.230771958300267, "replay_num_supervised_tokens": 74858, "total_loss": 1.5244140999811329, "replay_lambda": 1.0, "buffer_rows": 1216, "buffer_prompts": 226, "replay_draw_n_groups": 822, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.00039385584875935406, "replay_draw_max_q": 0.0027569909413154787}
+{"global_step": 20, "epoch": 0, "batch_id": 1279, "num_supervised_tokens": 36949, "learning_rate": 1e-05, "batch_loss": 0.28282994614458495, "token_entropy": 0.2925199919416216, "gradient_norm_pre_clip": 1.6472964798095169, "gradient_norm_post_clip": 1.6472964798095169, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 842.0, "replay_loss": 0.45857014651977457, "replay_num_supervised_tokens": 63621, "total_loss": 0.7414000926643596, "replay_lambda": 1.0, "buffer_rows": 1280, "buffer_prompts": 229, "replay_draw_n_groups": 865, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0003773584905660377, "replay_draw_max_q": 0.0026415094339622643, "parameter_l2_norm": 1111.3978625554496, "parameter_l2_from_init": 3.4349655863462, "relative_parameter_l2_from_init": 0.0030906599995468094, "update_norm": 1.5085309684975499, "relative_update_norm": 0.0013573275775688174}
+{"global_step": 21, "epoch": 0, "batch_id": 1343, "num_supervised_tokens": 36603, "learning_rate": 1e-05, "batch_loss": 0.28660133734365384, "token_entropy": 0.2875741907499157, "gradient_norm_pre_clip": 1.450909149489276, "gradient_norm_post_clip": 1.450909149489276, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2607.0, "replay_loss": 0.7505077701554744, "replay_num_supervised_tokens": 68151, "total_loss": 1.0371091074991283, "replay_lambda": 1.0, "buffer_rows": 1344, "buffer_prompts": 232, "replay_draw_n_groups": 909, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00035971223021582735, "replay_draw_max_q": 0.002517985611510791}
+{"global_step": 22, "epoch": 0, "batch_id": 1407, "num_supervised_tokens": 36508, "learning_rate": 1e-05, "batch_loss": 0.26270028079338587, "token_entropy": 0.26632453066450856, "gradient_norm_pre_clip": 1.6024488595431075, "gradient_norm_post_clip": 1.6024488595431075, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 937.0, "replay_loss": 0.3528687940390012, "replay_num_supervised_tokens": 67445, "total_loss": 0.615569074832387, "replay_lambda": 1.0, "buffer_rows": 1408, "buffer_prompts": 233, "replay_draw_n_groups": 957, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00034153005464480874, "replay_draw_max_q": 0.0023907103825136613}
+{"global_step": 23, "epoch": 0, "batch_id": 1471, "num_supervised_tokens": 34482, "learning_rate": 1e-05, "batch_loss": 0.2824196402969002, "token_entropy": 0.2843548912617856, "gradient_norm_pre_clip": 1.569191426768243, "gradient_norm_post_clip": 1.569191426768243, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2066.0, "replay_loss": 0.7436383602832984, "replay_num_supervised_tokens": 65547, "total_loss": 1.0260580005801987, "replay_lambda": 1.0, "buffer_rows": 1472, "buffer_prompts": 233, "replay_draw_n_groups": 1000, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0003290556103981573, "replay_draw_max_q": 0.002303389272787101}
+{"global_step": 24, "epoch": 0, "batch_id": 1535, "num_supervised_tokens": 34610, "learning_rate": 1e-05, "batch_loss": 0.27085319660901075, "token_entropy": 0.27897650821468656, "gradient_norm_pre_clip": 1.6671984028573155, "gradient_norm_post_clip": 1.6671984028573155, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2607.0, "replay_loss": 0.7687445299907036, "replay_num_supervised_tokens": 69251, "total_loss": 1.0395977265997143, "replay_lambda": 1.0, "buffer_rows": 1536, "buffer_prompts": 235, "replay_draw_n_groups": 1041, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0003194888178913738, "replay_draw_max_q": 0.0022364217252396168}
+{"global_step": 25, "epoch": 0, "batch_id": 1599, "num_supervised_tokens": 36993, "learning_rate": 1e-05, "batch_loss": 0.24958321657493496, "token_entropy": 0.2511808061124093, "gradient_norm_pre_clip": 1.5101344366370257, "gradient_norm_post_clip": 1.5101344366370257, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.5844965421268489, "replay_num_supervised_tokens": 63704, "total_loss": 0.8340797587017839, "replay_lambda": 1.0, "buffer_rows": 1600, "buffer_prompts": 235, "replay_draw_n_groups": 1087, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.0003059039461609055, "replay_draw_max_q": 0.0021413276231263384}
+{"global_step": 26, "epoch": 0, "batch_id": 1663, "num_supervised_tokens": 35574, "learning_rate": 1e-05, "batch_loss": 0.2485106409175347, "token_entropy": 0.24757321187709552, "gradient_norm_pre_clip": 1.6978308997593479, "gradient_norm_post_clip": 1.6978308997593479, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 909.0, "replay_loss": 0.345347086761468, "replay_num_supervised_tokens": 59788, "total_loss": 0.5938577276790027, "replay_lambda": 1.0, "buffer_rows": 1664, "buffer_prompts": 236, "replay_draw_n_groups": 1131, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0002957704821058858, "replay_draw_max_q": 0.002070393374741201}
+{"global_step": 27, "epoch": 0, "batch_id": 1727, "num_supervised_tokens": 31824, "learning_rate": 1e-05, "batch_loss": 0.25033938840416514, "token_entropy": 0.2442699487992062, "gradient_norm_pre_clip": 1.7518673529265885, "gradient_norm_post_clip": 1.7518673529265885, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1052.0, "replay_loss": 0.34690263079825956, "replay_num_supervised_tokens": 61331, "total_loss": 0.5972420192024247, "replay_lambda": 1.0, "buffer_rows": 1728, "buffer_prompts": 237, "replay_draw_n_groups": 1174, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00028677946659019213, "replay_draw_max_q": 0.002007456266131345}
+{"global_step": 28, "epoch": 0, "batch_id": 1791, "num_supervised_tokens": 33894, "learning_rate": 1e-05, "batch_loss": 0.239405210872194, "token_entropy": 0.23597142216223296, "gradient_norm_pre_clip": 1.5856931804024654, "gradient_norm_post_clip": 1.5856931804024654, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2683.0, "replay_loss": 0.6361599087237823, "replay_num_supervised_tokens": 58284, "total_loss": 0.8755651195959763, "replay_lambda": 1.0, "buffer_rows": 1792, "buffer_prompts": 242, "replay_draw_n_groups": 1214, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.0002788622420524261, "replay_draw_max_q": 0.0019520356943669827}
+{"global_step": 29, "epoch": 0, "batch_id": 1855, "num_supervised_tokens": 32475, "learning_rate": 1e-05, "batch_loss": 0.2324082772363783, "token_entropy": 0.2359147242310416, "gradient_norm_pre_clip": 1.7293557318911776, "gradient_norm_post_clip": 1.7293557318911776, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1052.0, "replay_loss": 0.33124546408003613, "replay_num_supervised_tokens": 59408, "total_loss": 0.5636537413164144, "replay_lambda": 1.0, "buffer_rows": 1856, "buffer_prompts": 243, "replay_draw_n_groups": 1258, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00026910656620021526, "replay_draw_max_q": 0.001883745963401507}
+{"global_step": 30, "epoch": 0, "batch_id": 1919, "num_supervised_tokens": 30711, "learning_rate": 1e-05, "batch_loss": 0.23915925339948277, "token_entropy": 0.24133993987674407, "gradient_norm_pre_clip": 1.7292416103739752, "gradient_norm_post_clip": 1.7292416103739752, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.6652637723216795, "replay_num_supervised_tokens": 62025, "total_loss": 0.9044230257211623, "replay_lambda": 1.0, "buffer_rows": 1920, "buffer_prompts": 243, "replay_draw_n_groups": 1302, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00026191723415400735, "replay_draw_max_q": 0.0018334206390780514, "parameter_l2_norm": 1111.396321292217, "parameter_l2_from_init": 4.241494224992014, "relative_parameter_l2_from_init": 0.0038163458148166704, "update_norm": 1.2325183426307378, "relative_update_norm": 0.001108981844746159}
+{"global_step": 31, "epoch": 0, "batch_id": 1983, "num_supervised_tokens": 31193, "learning_rate": 1e-05, "batch_loss": 0.20339613180867072, "token_entropy": 0.21261428935978835, "gradient_norm_pre_clip": 1.5468327919035931, "gradient_norm_post_clip": 1.5468327919035931, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1052.0, "replay_loss": 0.31275814335953567, "replay_num_supervised_tokens": 56608, "total_loss": 0.5161542751682064, "replay_lambda": 1.0, "buffer_rows": 1984, "buffer_prompts": 245, "replay_draw_n_groups": 1342, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00025510204081632655, "replay_draw_max_q": 0.0017857142857142857}
+{"global_step": 32, "epoch": 0, "batch_id": 2047, "num_supervised_tokens": 29719, "learning_rate": 1e-05, "batch_loss": 0.20308903258515362, "token_entropy": 0.20920441470679993, "gradient_norm_pre_clip": 1.6701539348490528, "gradient_norm_post_clip": 1.6701539348490528, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 909.0, "replay_loss": 0.3261531246098852, "replay_num_supervised_tokens": 58912, "total_loss": 0.5292421571950389, "replay_lambda": 1.0, "buffer_rows": 2048, "buffer_prompts": 246, "replay_draw_n_groups": 1387, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.00024752475247524753, "replay_draw_max_q": 0.0017326732673267327}
+{"global_step": 33, "epoch": 0, "batch_id": 2111, "num_supervised_tokens": 30714, "learning_rate": 1e-05, "batch_loss": 0.19483429345606054, "token_entropy": 0.1986225495435006, "gradient_norm_pre_clip": 1.6562720803949567, "gradient_norm_post_clip": 1.6562720803949567, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.49438598301813497, "replay_num_supervised_tokens": 57098, "total_loss": 0.6892202764741955, "replay_lambda": 1.0, "buffer_rows": 2112, "buffer_prompts": 247, "replay_draw_n_groups": 1432, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.00024073182474723158, "replay_draw_max_q": 0.001685122773230621}
+{"global_step": 34, "epoch": 0, "batch_id": 2175, "num_supervised_tokens": 29442, "learning_rate": 1e-05, "batch_loss": 0.17378470559419923, "token_entropy": 0.17981824077396044, "gradient_norm_pre_clip": 1.7964181806134365, "gradient_norm_post_clip": 1.7964181806134365, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.6407072470683107, "replay_num_supervised_tokens": 56886, "total_loss": 0.81449195266251, "replay_lambda": 1.0, "buffer_rows": 2176, "buffer_prompts": 248, "replay_draw_n_groups": 1475, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.000234192037470726, "replay_draw_max_q": 0.001639344262295082}
+{"global_step": 35, "epoch": 0, "batch_id": 2239, "num_supervised_tokens": 34180, "learning_rate": 1e-05, "batch_loss": 0.15582537771550847, "token_entropy": 0.16569357747010421, "gradient_norm_pre_clip": 1.4994683160396682, "gradient_norm_post_clip": 1.4994683160396682, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 994.0, "replay_loss": 0.3020917675797046, "replay_num_supervised_tokens": 54454, "total_loss": 0.45791714529521305, "replay_lambda": 1.0, "buffer_rows": 2240, "buffer_prompts": 250, "replay_draw_n_groups": 1515, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00022935779816513763, "replay_draw_max_q": 0.0016055045871559634}
+{"global_step": 36, "epoch": 0, "batch_id": 2303, "num_supervised_tokens": 30222, "learning_rate": 1e-05, "batch_loss": 0.16117358080604932, "token_entropy": 0.17326267628167688, "gradient_norm_pre_clip": 1.663196566607447, "gradient_norm_post_clip": 1.663196566607447, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 891.0, "replay_loss": 0.3652856948556256, "replay_num_supervised_tokens": 55442, "total_loss": 0.5264592756616749, "replay_lambda": 1.0, "buffer_rows": 2304, "buffer_prompts": 254, "replay_draw_n_groups": 1560, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0002228163992869875, "replay_draw_max_q": 0.0015597147950089127}
+{"global_step": 37, "epoch": 0, "batch_id": 2367, "num_supervised_tokens": 31163, "learning_rate": 1e-05, "batch_loss": 0.15561207130716784, "token_entropy": 0.16441923304347467, "gradient_norm_pre_clip": 1.5114752028703107, "gradient_norm_post_clip": 1.5114752028703107, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2750.0, "replay_loss": 0.8392734560322773, "replay_num_supervised_tokens": 56514, "total_loss": 0.9948855273394451, "replay_lambda": 1.0, "buffer_rows": 2368, "buffer_prompts": 256, "replay_draw_n_groups": 1603, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0002173913043478261, "replay_draw_max_q": 0.0015217391304347826}
+{"global_step": 38, "epoch": 0, "batch_id": 2431, "num_supervised_tokens": 28931, "learning_rate": 1e-05, "batch_loss": 0.1558054896792442, "token_entropy": 0.16686111165520348, "gradient_norm_pre_clip": 1.5417049182839442, "gradient_norm_post_clip": 1.5417049182839442, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.7945320496294214, "replay_num_supervised_tokens": 55504, "total_loss": 0.9503375393086656, "replay_lambda": 1.0, "buffer_rows": 2432, "buffer_prompts": 256, "replay_draw_n_groups": 1649, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00021128248468201986, "replay_draw_max_q": 0.001478977392774139}
+{"global_step": 39, "epoch": 0, "batch_id": 2495, "num_supervised_tokens": 29457, "learning_rate": 1e-05, "batch_loss": 0.13692230397064126, "token_entropy": 0.14898236430735007, "gradient_norm_pre_clip": 1.602050962613636, "gradient_norm_post_clip": 1.602050962613636, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 928.0, "replay_loss": 0.27585848674867697, "replay_num_supervised_tokens": 55298, "total_loss": 0.4127807907193182, "replay_lambda": 1.0, "buffer_rows": 2496, "buffer_prompts": 257, "replay_draw_n_groups": 1692, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0002063983488132095, "replay_draw_max_q": 0.0014447884416924665}
+{"global_step": 40, "epoch": 0, "batch_id": 2559, "num_supervised_tokens": 30001, "learning_rate": 1e-05, "batch_loss": 0.13099577556851646, "token_entropy": 0.1385600525936855, "gradient_norm_pre_clip": 1.5116321859425217, "gradient_norm_post_clip": 1.5116321859425217, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1101.0, "replay_loss": 0.2804836460824546, "replay_num_supervised_tokens": 54996, "total_loss": 0.41147942165097107, "replay_lambda": 1.0, "buffer_rows": 2560, "buffer_prompts": 258, "replay_draw_n_groups": 1731, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00020308692120227456, "replay_draw_max_q": 0.001421608448415922, "parameter_l2_norm": 1111.3941106838272, "parameter_l2_from_init": 4.865145070999891, "relative_parameter_l2_from_init": 0.004377484689424833, "update_norm": 1.1078899603632169, "relative_update_norm": 0.0009968470677620793}
+{"global_step": 41, "epoch": 0, "batch_id": 2623, "num_supervised_tokens": 29105, "learning_rate": 1e-05, "batch_loss": 0.14069254274254933, "token_entropy": 0.13531312714776433, "gradient_norm_pre_clip": 1.8001802225562935, "gradient_norm_post_clip": 1.8001802225562935, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 875.0, "replay_loss": 0.3464121310845347, "replay_num_supervised_tokens": 53964, "total_loss": 0.48710467382708406, "replay_lambda": 1.0, "buffer_rows": 2624, "buffer_prompts": 258, "replay_draw_n_groups": 1773, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00019952114924181964, "replay_draw_max_q": 0.0013966480446927375}
+{"global_step": 42, "epoch": 0, "batch_id": 2687, "num_supervised_tokens": 28971, "learning_rate": 1e-05, "batch_loss": 0.11598968091867508, "token_entropy": 0.1258044087320205, "gradient_norm_pre_clip": 1.7192420774652748, "gradient_norm_post_clip": 1.7192420774652748, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 885.0, "replay_loss": 0.27579555735234185, "replay_num_supervised_tokens": 52305, "total_loss": 0.3917852382710169, "replay_lambda": 1.0, "buffer_rows": 2688, "buffer_prompts": 259, "replay_draw_n_groups": 1814, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00019638648860958367, "replay_draw_max_q": 0.0013747054202670856}
+{"global_step": 43, "epoch": 0, "batch_id": 2751, "num_supervised_tokens": 28111, "learning_rate": 1e-05, "batch_loss": 0.11640987586141119, "token_entropy": 0.123394567469201, "gradient_norm_pre_clip": 1.721464434934602, "gradient_norm_post_clip": 1.721464434934602, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2238.0, "replay_loss": 0.7789977253647521, "replay_num_supervised_tokens": 52160, "total_loss": 0.8954076012261633, "replay_lambda": 1.0, "buffer_rows": 2752, "buffer_prompts": 259, "replay_draw_n_groups": 1852, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00019372336303758234, "replay_draw_max_q": 0.0013560635412630763}
+{"global_step": 44, "epoch": 0, "batch_id": 2815, "num_supervised_tokens": 29323, "learning_rate": 1e-05, "batch_loss": 0.11735777733594366, "token_entropy": 0.12337692352327184, "gradient_norm_pre_clip": 1.7326899286716126, "gradient_norm_post_clip": 1.7326899286716126, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 906.0, "replay_loss": 0.28322091618381023, "replay_num_supervised_tokens": 52079, "total_loss": 0.4005786935197539, "replay_lambda": 1.0, "buffer_rows": 2816, "buffer_prompts": 259, "replay_draw_n_groups": 1893, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00019094901661256445, "replay_draw_max_q": 0.001336643116287951}
+{"global_step": 45, "epoch": 0, "batch_id": 2879, "num_supervised_tokens": 26914, "learning_rate": 1e-05, "batch_loss": 0.11827225544776923, "token_entropy": 0.12518191537788315, "gradient_norm_pre_clip": 1.8051009768699908, "gradient_norm_post_clip": 1.8051009768699908, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 860.0, "replay_loss": 0.26397500846016503, "replay_num_supervised_tokens": 50663, "total_loss": 0.3822472639079343, "replay_lambda": 1.0, "buffer_rows": 2880, "buffer_prompts": 259, "replay_draw_n_groups": 1932, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.00018839487565938207, "replay_draw_max_q": 0.0013187641296156744}
+{"global_step": 46, "epoch": 0, "batch_id": 2943, "num_supervised_tokens": 25515, "learning_rate": 1e-05, "batch_loss": 0.11889402320637332, "token_entropy": 0.12615114989851633, "gradient_norm_pre_clip": 1.8130480727412657, "gradient_norm_post_clip": 1.8130480727412657, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 831.0, "replay_loss": 0.2743083260963801, "replay_num_supervised_tokens": 50629, "total_loss": 0.39320234930275344, "replay_lambda": 1.0, "buffer_rows": 2944, "buffer_prompts": 259, "replay_draw_n_groups": 1976, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.00018511662347278786, "replay_draw_max_q": 0.001295816364309515}
+{"global_step": 47, "epoch": 0, "batch_id": 3007, "num_supervised_tokens": 30018, "learning_rate": 1e-05, "batch_loss": 0.09130447192825077, "token_entropy": 0.09758351671837881, "gradient_norm_pre_clip": 1.6390507105357708, "gradient_norm_post_clip": 1.6390507105357708, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.6512910097399425, "replay_num_supervised_tokens": 51641, "total_loss": 0.7425954816681932, "replay_lambda": 1.0, "buffer_rows": 3008, "buffer_prompts": 261, "replay_draw_n_groups": 2015, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00018291567587342235, "replay_draw_max_q": 0.0012804097311139564}
+{"global_step": 48, "epoch": 0, "batch_id": 3071, "num_supervised_tokens": 30651, "learning_rate": 1e-05, "batch_loss": 0.09363910843832153, "token_entropy": 0.10157329741902953, "gradient_norm_pre_clip": 1.7384223754289354, "gradient_norm_post_clip": 1.7384223754289354, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 828.0, "replay_loss": 0.2548917831483767, "replay_num_supervised_tokens": 48394, "total_loss": 0.3485308915866982, "replay_lambda": 1.0, "buffer_rows": 3072, "buffer_prompts": 261, "replay_draw_n_groups": 2055, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.00018040772145047808, "replay_draw_max_q": 0.0012628540501533466}
+{"global_step": 49, "epoch": 0, "batch_id": 3135, "num_supervised_tokens": 28413, "learning_rate": 1e-05, "batch_loss": 0.11846688657489536, "token_entropy": 0.11176109901347399, "gradient_norm_pre_clip": 1.9442035190040354, "gradient_norm_post_clip": 1.9442035190040354, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 909.0, "replay_loss": 0.2930581772129518, "replay_num_supervised_tokens": 49156, "total_loss": 0.41152506378784715, "replay_lambda": 1.0, "buffer_rows": 3136, "buffer_prompts": 261, "replay_draw_n_groups": 2095, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.00017822135091783995, "replay_draw_max_q": 0.0012475494564248798}
+{"global_step": 50, "epoch": 0, "batch_id": 3199, "num_supervised_tokens": 27750, "learning_rate": 1e-05, "batch_loss": 0.09293718470223583, "token_entropy": 0.10572000146126961, "gradient_norm_pre_clip": 1.7996429835129268, "gradient_norm_post_clip": 1.7996429835129268, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2057.0, "replay_loss": 0.40876172564750507, "replay_num_supervised_tokens": 50443, "total_loss": 0.5016989103497409, "replay_lambda": 1.0, "buffer_rows": 3200, "buffer_prompts": 261, "replay_draw_n_groups": 2132, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0001765848490199541, "replay_draw_max_q": 0.0012360939431396785, "parameter_l2_norm": 1111.3915964272587, "parameter_l2_from_init": 5.389663604053794, "relative_parameter_l2_from_init": 0.004849427830740276, "update_norm": 1.0658445285968565, "relative_update_norm": 0.0009590179843208997}
+{"global_step": 51, "epoch": 0, "batch_id": 3263, "num_supervised_tokens": 24101, "learning_rate": 1e-05, "batch_loss": 0.09969912276452847, "token_entropy": 0.10507036540464842, "gradient_norm_pre_clip": 1.7002466841947295, "gradient_norm_post_clip": 1.7002466841947295, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1134.0, "replay_loss": 0.35149079907869263, "replay_num_supervised_tokens": 52369, "total_loss": 0.4511899218432211, "replay_lambda": 1.0, "buffer_rows": 3264, "buffer_prompts": 261, "replay_draw_n_groups": 2172, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00017464198393293747, "replay_draw_max_q": 0.0012224938875305623}
+{"global_step": 52, "epoch": 0, "batch_id": 3327, "num_supervised_tokens": 29272, "learning_rate": 1e-05, "batch_loss": 0.10235376243616232, "token_entropy": 0.10182479666568894, "gradient_norm_pre_clip": 1.8529837033953105, "gradient_norm_post_clip": 1.8529837033953105, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1030.0, "replay_loss": 0.27888920585836685, "replay_num_supervised_tokens": 48876, "total_loss": 0.3812429682945292, "replay_lambda": 1.0, "buffer_rows": 3328, "buffer_prompts": 262, "replay_draw_n_groups": 2212, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.00017277125086385625, "replay_draw_max_q": 0.0012093987560469939}
+{"global_step": 53, "epoch": 0, "batch_id": 3391, "num_supervised_tokens": 29206, "learning_rate": 1e-05, "batch_loss": 0.120545913663601, "token_entropy": 0.10180698276700806, "gradient_norm_pre_clip": 3.0536403490663244, "gradient_norm_post_clip": 3.0536403490663244, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1065.0, "replay_loss": 0.4536198954706826, "replay_num_supervised_tokens": 52282, "total_loss": 0.5741658091342836, "replay_lambda": 1.0, "buffer_rows": 3392, "buffer_prompts": 263, "replay_draw_n_groups": 2251, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00017085255424568598, "replay_draw_max_q": 0.0011959678797198018}
+{"global_step": 54, "epoch": 0, "batch_id": 3455, "num_supervised_tokens": 28302, "learning_rate": 1e-05, "batch_loss": 0.10053329899973978, "token_entropy": 0.12018374171528949, "gradient_norm_pre_clip": 1.7677232197512995, "gradient_norm_post_clip": 1.7677232197512995, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1174.0, "replay_loss": 0.3364785349123233, "replay_num_supervised_tokens": 47877, "total_loss": 0.4370118339120631, "replay_lambda": 1.0, "buffer_rows": 3456, "buffer_prompts": 263, "replay_draw_n_groups": 2288, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00016957775139901646, "replay_draw_max_q": 0.0011870442597931151}
+{"global_step": 55, "epoch": 0, "batch_id": 3519, "num_supervised_tokens": 24971, "learning_rate": 1e-05, "batch_loss": 0.10312276328580253, "token_entropy": 0.1327300648977423, "gradient_norm_pre_clip": 2.1290132199573764, "gradient_norm_post_clip": 2.1290132199573764, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 855.0, "replay_loss": 0.27314799211385143, "replay_num_supervised_tokens": 47786, "total_loss": 0.37627075539965393, "replay_lambda": 1.0, "buffer_rows": 3520, "buffer_prompts": 264, "replay_draw_n_groups": 2325, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0001682935038707506, "replay_draw_max_q": 0.0011780545270952542}
+{"global_step": 56, "epoch": 0, "batch_id": 3583, "num_supervised_tokens": 31094, "learning_rate": 1e-05, "batch_loss": 0.0798841936771008, "token_entropy": 0.0977979728613834, "gradient_norm_pre_clip": 1.6578445305136202, "gradient_norm_post_clip": 1.6578445305136202, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 844.0, "replay_loss": 0.24163021098585655, "replay_num_supervised_tokens": 47357, "total_loss": 0.32151440466295733, "replay_lambda": 1.0, "buffer_rows": 3584, "buffer_prompts": 264, "replay_draw_n_groups": 2364, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00016641704110500916, "replay_draw_max_q": 0.001164919287735064}
+{"global_step": 57, "epoch": 0, "batch_id": 3647, "num_supervised_tokens": 28839, "learning_rate": 1e-05, "batch_loss": 0.0829888746295977, "token_entropy": 0.10330412663063909, "gradient_norm_pre_clip": 1.7510905715281582, "gradient_norm_post_clip": 1.7510905715281582, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1030.0, "replay_loss": 0.26268773482021657, "replay_num_supervised_tokens": 49085, "total_loss": 0.34567660944981426, "replay_lambda": 1.0, "buffer_rows": 3648, "buffer_prompts": 264, "replay_draw_n_groups": 2404, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.00016501650165016502, "replay_draw_max_q": 0.0011551155115511551}
+{"global_step": 58, "epoch": 0, "batch_id": 3711, "num_supervised_tokens": 23554, "learning_rate": 1e-05, "batch_loss": 0.09490526056955778, "token_entropy": 0.10653261552493075, "gradient_norm_pre_clip": 1.965334709904512, "gradient_norm_post_clip": 1.965334709904512, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 1083.0, "replay_loss": 0.2881009713870823, "replay_num_supervised_tokens": 50684, "total_loss": 0.3830062319566401, "replay_lambda": 1.0, "buffer_rows": 3712, "buffer_prompts": 267, "replay_draw_n_groups": 2439, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00016377333770062233, "replay_draw_max_q": 0.0011464133639043563}
+{"global_step": 59, "epoch": 0, "batch_id": 3775, "num_supervised_tokens": 27043, "learning_rate": 1e-05, "batch_loss": 0.0938528547992554, "token_entropy": 0.10071301842010907, "gradient_norm_pre_clip": 2.1026554120490997, "gradient_norm_post_clip": 2.1026554120490997, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 3030.0, "replay_loss": 0.8507074540349401, "replay_num_supervised_tokens": 52551, "total_loss": 0.9445603088341955, "replay_lambda": 1.0, "buffer_rows": 3776, "buffer_prompts": 268, "replay_draw_n_groups": 2480, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00016215339711366953, "replay_draw_max_q": 0.0011350737797956867}
+{"global_step": 60, "epoch": 0, "batch_id": 3839, "num_supervised_tokens": 29023, "learning_rate": 1e-05, "batch_loss": 0.07872022673846035, "token_entropy": 0.0868048182083549, "gradient_norm_pre_clip": 1.8053390269962566, "gradient_norm_post_clip": 1.8053390269962566, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.8087132914250978, "replay_num_supervised_tokens": 51890, "total_loss": 0.8874335181635582, "replay_lambda": 1.0, "buffer_rows": 3840, "buffer_prompts": 268, "replay_draw_n_groups": 2515, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00016090104585679806, "replay_draw_max_q": 0.0011263073209975866, "parameter_l2_norm": 1111.3894946515686, "parameter_l2_from_init": 5.841757933768431, "relative_parameter_l2_from_init": 0.005256206246927334, "update_norm": 1.088628955451162, "relative_update_norm": 0.000979520645723269}
+{"global_step": 61, "epoch": 0, "batch_id": 3903, "num_supervised_tokens": 29055, "learning_rate": 1e-05, "batch_loss": 0.09705099902349594, "token_entropy": 0.09783802925258224, "gradient_norm_pre_clip": 1.8509719375053222, "gradient_norm_post_clip": 1.8509719375053222, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 815.0, "replay_loss": 0.2542322547815211, "replay_num_supervised_tokens": 50613, "total_loss": 0.35128325380501707, "replay_lambda": 1.0, "buffer_rows": 3904, "buffer_prompts": 269, "replay_draw_n_groups": 2554, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00015959144589849984, "replay_draw_max_q": 0.0011171401212894988}
+{"global_step": 62, "epoch": 0, "batch_id": 3967, "num_supervised_tokens": 27553, "learning_rate": 1e-05, "batch_loss": 0.09067884723760697, "token_entropy": 0.09682590470912647, "gradient_norm_pre_clip": 1.7414810598328423, "gradient_norm_post_clip": 1.7414810598328423, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2017.0, "replay_loss": 0.6585621761882817, "replay_num_supervised_tokens": 53712, "total_loss": 0.7492410234258887, "replay_lambda": 1.0, "buffer_rows": 3968, "buffer_prompts": 270, "replay_draw_n_groups": 2593, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00015815277558121145, "replay_draw_max_q": 0.0011070694290684802}
+{"global_step": 63, "epoch": 0, "batch_id": 4031, "num_supervised_tokens": 29373, "learning_rate": 1e-05, "batch_loss": 0.08745401559352999, "token_entropy": 0.0889444411910924, "gradient_norm_pre_clip": 1.7122687681604474, "gradient_norm_post_clip": 1.7122687681604474, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 850.0, "replay_loss": 0.2661914228981543, "replay_num_supervised_tokens": 44273, "total_loss": 0.3536454384916843, "replay_lambda": 1.0, "buffer_rows": 4032, "buffer_prompts": 270, "replay_draw_n_groups": 2631, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00015681354869060687, "replay_draw_max_q": 0.0010976948408342481}
+{"global_step": 64, "epoch": 0, "batch_id": 4095, "num_supervised_tokens": 30287, "learning_rate": 1e-05, "batch_loss": 0.09616264606213684, "token_entropy": 0.09707552578944632, "gradient_norm_pre_clip": 1.752515525931961, "gradient_norm_post_clip": 1.752515525931961, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 2871.0, "replay_loss": 0.6710210170893818, "replay_num_supervised_tokens": 49699, "total_loss": 0.7671836631515186, "replay_lambda": 1.0, "buffer_rows": 4096, "buffer_prompts": 270, "replay_draw_n_groups": 2665, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00015583606046439146, "replay_draw_max_q": 0.0010908524232507402}
+{"global_step": 65, "epoch": 0, "batch_id": 4105, "num_supervised_tokens": 4748, "learning_rate": 1e-05, "batch_loss": 0.21864656124753593, "token_entropy": 0.20604007467062613, "gradient_norm_pre_clip": 3.6809591849601584, "gradient_norm_post_clip": 3.6809591849601584, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 73.154296875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 73.154296875, "max_sequence_length": 704.0, "replay_loss": 0.37019532127472515, "replay_num_supervised_tokens": 7824, "total_loss": 0.5888418825222611, "replay_lambda": 1.0, "buffer_rows": 4106, "buffer_prompts": 270, "replay_draw_n_groups": 2705, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00015422578655151142, "replay_draw_max_q": 0.00107958050586058, "parameter_l2_norm": 1111.388887786505, "parameter_l2_from_init": 6.079890918161503, "relative_parameter_l2_from_init": 0.005470469846062616, "update_norm": 0.6101580479163062, "relative_update_norm": 0.0005490049924212632}

training/llama3b_sft_rp_hard_lam1/provenance.json

@@ -0,0 +1,25 @@
+{
+  "base_model": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "dataset": "llama3b_code_sft_data",
+  "dataset_sha256": "69ec8901893709bcba86b78d434c5995f523c783916edffa0129aed8649040de",
+  "durable_jsonl_sha256": "757270437abe9b280a7104cb5b13223f32d782444e42bda77e5907c4c1b5d7e5",
+  "effective_batch_size": 256,
+  "max_seq_length": 4096,
+  "metrics_log_every": 10,
+  "micro_batch_size": 1,
+  "model_name": "llama3b",
+  "num_checkpoints": 10,
+  "order": "ordered",
+  "replay_lambda": 1.0,
+  "replay_strategy": "hard",
+  "resume_generation": 1,
+  "resume_helper_sha256": "e30bddb7e051f29d4a16700b2b5dcec4e572785be673ab7d42480c0c4f3f0707",
+  "resume_step": 0,
+  "run_sft_sha256": "f4ee1ff82e7da742db079178b8d3695d2142f1db42096b0f1dcb5d837baea263",
+  "run_sweep_sha256": "aec506941a4ca39773d465ce72586294a32e9b9d1761902201976fa8d28386fc",
+  "save_optimizer_every_n_checkpoints": 5,
+  "save_optimizer_first_checkpoint": true,
+  "schema_version": 1,
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "world_size": 4
+}

training/llama3b_sft_rp_hard_lam1/sft_config.json

@@ -0,0 +1,56 @@
+{
+  "generation_logs_dir": null,
+  "dataset": "llama3b_code_sft_data",
+  "positives_file": "/mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl",
+  "model_path": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "output_dir": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922",
+  "lr": 1e-05,
+  "batch_size": 1,
+  "effective_batch_size": 256,
+  "weight_decay": 0.01,
+  "warmup_ratio": 0.0,
+  "schedule": "constant",
+  "max_seq_length": 4096,
+  "seed": 42,
+  "bf16": true,
+  "attn_impl": "auto",
+  "precision": "mixed_bf16",
+  "grad_clip": 0.0,
+  "loss_agg_mode": "token-mean",
+  "ordered": true,
+  "within_step_shuffle": false,
+  "system_prompt_mode": "from_record",
+  "score_threshold": 0.0,
+  "drop_all_pass_groups": false,
+  "max_rl_step": null,
+  "num_epochs": 1,
+  "replay_strategy": "hard",
+  "replay_lambda": 1.0,
+  "replay_batch_size": 1,
+  "replay_cooldown_step": 3.0,
+  "replay_ce_tau": 1.0,
+  "replay_ce_init": "insert",
+  "replay_group_size": 8,
+  "replay_seed": null,
+  "max_train_steps": null,
+  "num_checkpoints": 10,
+  "checkpoint_after_resume_steps": null,
+  "save_optimizer_state": true,
+  "prune_old_optimizer": true,
+  "save_optimizer_every_n_ckpts": 5,
+  "save_optimizer_first_ckpt": true,
+  "snapshot_dtype": "float32",
+  "snapshot_device": "cpu",
+  "metrics_log_every": 10,
+  "ddp_timeout_seconds": 3600,
+  "memory_canary_longest_first": false,
+  "log_cuda_memory": true,
+  "save_final_model": true,
+  "wandb_project": "rl_forgetting_sft",
+  "wandb_entity": "jingyanshen-new-york-university",
+  "wandb_run_name": "llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922_segment_g1",
+  "wandb_group": "code_sft_20260922_llama3b",
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "resume_from": null,
+  "grad_accum_steps": 64
+}

training/llama3b_sft_rp_hard_lam1/training_complete.json

@@ -0,0 +1,7 @@
+{
+  "azure_run_id": "yellow_arm_6lw7gf1fvy",
+  "completed_at": "2026-09-23T02:19:36.342716+00:00",
+  "host": "node-0",
+  "schema_version": 1,
+  "status": "training_complete"
+}

training/llama3b_sft_rp_hard_lam1/wandb/run-20260923_014857-79qp1zp0/files/config.yaml

@@ -0,0 +1,253 @@
+_wandb:
+    value:
+        cli_version: 0.28.0
+        e:
+            4gm1jbz4pougcac6c2xr1sw2b6quaw96:
+                args:
+                    - --dataset
+                    - llama3b_code_sft_data
+                    - --replay_strategy
+                    - hard
+                    - --replay_lambda
+                    - "1.0"
+                    - --replay_cooldown_step
+                    - "3"
+                    - --replay_ce_tau
+                    - "1.0"
+                    - --replay_ce_init
+                    - insert
+                    - --replay_group_size
+                    - "8"
+                    - --model_path
+                    - /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+                    - --output_dir
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922
+                    - --ordered
+                    - --system_prompt_mode
+                    - from_record
+                    - --lr
+                    - "1e-5"
+                    - --batch_size
+                    - "1"
+                    - --effective_batch_size
+                    - "256"
+                    - --max_seq_length
+                    - "4096"
+                    - --ddp_timeout_seconds
+                    - "3600"
+                    - --schedule
+                    - constant
+                    - --warmup_ratio
+                    - "0"
+                    - --num_epochs
+                    - "1"
+                    - --seed
+                    - "42"
+                    - --log_cuda_memory
+                    - --num_checkpoints
+                    - "10"
+                    - --save_optimizer_every_n_ckpts
+                    - "5"
+                    - --prune_old_optimizer
+                    - --save_optimizer_first_ckpt
+                    - --metrics_log_every
+                    - "10"
+                    - --precision
+                    - mixed_bf16
+                    - --grad_clip
+                    - "0"
+                    - --snapshot_device
+                    - cpu
+                    - --snapshot_dtype
+                    - float32
+                    - --wandb_project
+                    - rl_forgetting_sft
+                    - --wandb_entity
+                    - jingyanshen-new-york-university
+                    - --wandb_group
+                    - code_sft_20260922_llama3b
+                    - --wandb_run_name
+                    - llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922_segment_g1
+                    - --wandb_run_id_file
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+                codePath: sft/run_sft.py
+                codePathLocal: sft/run_sft.py
+                cpu_count: 96
+                cpu_count_logical: 96
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "30601604947968"
+                        used: "578502160384"
+                email: js15262@nyu.edu
+                executable: /tmp/rlf-code-runtime/bin/python
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-3b56a237-fc29-56a1-d432-428ca67ad7e4
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-6056d837-d995-d622-77fc-64519b1b237a
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-5099ed2a-4aa9-e0a3-4b19-55d9f18ca1f6
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-947282df-cba4-b58d-0319-dc7d31bb3cc9
+                host: node-0
+                memory:
+                    total: "2008079052800"
+                os: Linux-6.8.0-1046-azure-x86_64-with-glibc2.35
+                program: /scratch/amlt_code/sft/run_sft.py
+                python: CPython 3.10.21
+                root: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922
+                startedAt: "2026-09-23T01:48:57.607753Z"
+                writerId: 4gm1jbz4pougcac6c2xr1sw2b6quaw96
+        m: []
+        python_version: 3.10.21
+        t:
+            "1":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "2":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "3":
+                - 2
+                - 13
+                - 14
+                - 16
+                - 61
+            "4": 3.10.21
+            "5": 0.28.0
+            "6": 4.57.6
+            "12": 0.28.0
+            "13": linux-x86_64
+attn_impl:
+    value: auto
+batch_size:
+    value: 1
+bf16:
+    value: true
+checkpoint_after_resume_steps:
+    value: null
+dataset:
+    value: llama3b_code_sft_data
+ddp_timeout_seconds:
+    value: 3600
+drop_all_pass_groups:
+    value: false
+effective_batch_size:
+    value: 256
+generation_logs_dir:
+    value: null
+grad_accum_steps:
+    value: 64
+grad_clip:
+    value: 0
+log_cuda_memory:
+    value: true
+loss_agg_mode:
+    value: token-mean
+lr:
+    value: 1e-05
+max_rl_step:
+    value: null
+max_seq_length:
+    value: 4096
+max_train_steps:
+    value: null
+memory_canary_longest_first:
+    value: false
+metrics_log_every:
+    value: 10
+model_path:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+num_checkpoints:
+    value: 10
+num_epochs:
+    value: 1
+ordered:
+    value: true
+output_dir:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922
+positives_file:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl
+precision:
+    value: mixed_bf16
+prune_old_optimizer:
+    value: true
+replay_batch_size:
+    value: 1
+replay_ce_init:
+    value: insert
+replay_ce_tau:
+    value: 1
+replay_cooldown_step:
+    value: 3
+replay_group_size:
+    value: 8
+replay_lambda:
+    value: 1
+replay_seed:
+    value: null
+replay_strategy:
+    value: hard
+resume_from:
+    value: null
+save_final_model:
+    value: true
+save_optimizer_every_n_ckpts:
+    value: 5
+save_optimizer_first_ckpt:
+    value: true
+save_optimizer_state:
+    value: true
+schedule:
+    value: constant
+score_threshold:
+    value: 0
+seed:
+    value: 42
+snapshot_device:
+    value: cpu
+snapshot_dtype:
+    value: float32
+system_prompt_mode:
+    value: from_record
+wandb_entity:
+    value: jingyanshen-new-york-university
+wandb_group:
+    value: code_sft_20260922_llama3b
+wandb_project:
+    value: rl_forgetting_sft
+wandb_run_id_file:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+wandb_run_name:
+    value: llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922_segment_g1
+warmup_ratio:
+    value: 0
+weight_decay:
+    value: 0.01
+within_step_shuffle:
+    value: false

training/llama3b_sft_rp_hard_lam1/wandb/run-20260923_014857-79qp1zp0/files/requirements.txt

@@ -0,0 +1,266 @@
+colorful==0.5.8
+torchaudio==2.6.0
+referencing==0.37.0
+pyasn1_modules==0.4.2
+distlib==0.4.3
+frozenlist==1.8.0
+fastar==0.12.0
+nvidia-cusparselt-cu12==0.6.2
+python-json-logger==4.2.0
+websockets==16.1.1
+requests==2.34.2
+yarl==1.24.5
+pyparsing==3.3.2
+pydantic-settings==2.15.0
+safetensors==0.8.0
+cuda-pathfinder==1.8.1
+nvidia-cuda-runtime-cu12==12.4.127
+httpx==0.28.1
+gunicorn==26.2.0
+xgrammar==0.1.18
+hf-xet==1.6.0
+agent-detector==2.0.0
+hydra-core==1.3.6
+python-discovery==1.6.0
+ray==2.47.1
+kiwisolver==1.5.1
+cachetools==7.1.8
+pandas==2.3.3
+openai==3.13.0
+sqlparse==0.6.0
+lm-format-enforcer==0.10.12
+arckit==0.1.0
+Pygments==2.21.0
+jiter==0.16.0
+orjson==3.12.0
+skops==0.14.0
+depyf==0.18.0
+detect-installer==0.2.1
+accelerate==1.15.0
+pyfiglet==1.0.2
+GitPython==3.1.62
+sniffio==1.3.1
+numpy==2.2.6
+nvidia-cusparse-cu12==12.3.1.170
+huggingface_hub==0.36.2
+prometheus-fastapi-instrumentator==8.1.0
+einops==0.8.2
+gitdb==4.0.12
+pyarrow==25.0.1
+opencv-python-headless==5.0.0.93
+setuptools==80.9.0
+mistral_common==1.11.7
+diskcache==5.6.3
+interegular==0.3.3
+flash_attn==2.7.4.post1
+opentelemetry-exporter-otlp-proto-http==1.26.0
+zipp==4.1.0
+pybind11==3.1.0
+verl==0.4.1.dev0
+opentelemetry-exporter-prometheus==0.47b0
+llvmlite==0.44.0
+jsonschema-specifications==2025.9.1
+google-api-core==2.30.3
+msgpack==1.2.2
+ninja==1.13.2
+jsonschema==4.26.0
+nvidia-cuda-nvrtc-cu12==12.4.127
+pydantic==2.13.5
+rpds-py==0.30.0
+aiohttp-cors==0.8.1
+torchdata==0.11.0
+MarkupSafe==3.0.3
+antlr4-python3-runtime==4.9.3
+math-verify==0.9.0
+markdown-it-py==4.2.0
+nvidia-cublas-cu12==12.4.5.8
+mpmath==1.3.0
+python-dateutil==2.9.0.post0
+huey==3.4.0
+cupy-cuda12x==14.2.0
+psutil==7.2.2
+click==8.5.0
+astor==0.8.1
+uvloop==0.22.1
+tqdm==4.70.1
+scipy==1.15.3
+alembic==1.19.2
+SQLAlchemy==2.0.52
+tiktoken==0.14.0
+rich==15.0.0
+triton==3.2.0
+nvidia-cusolver-cu12==11.6.1.9
+outlines_core==0.1.26
+aiosignal==1.4.0
+exceptiongroup==1.3.1
+httptools==0.8.0
+pylatexenc==2.11
+Flask==3.1.3
+scikit-learn==1.7.2
+propcache==0.5.2
+PyYAML==6.0.3
+cloudpickle==3.1.2
+blinker==1.9.0
+protobuf==4.25.9
+pycountry==26.2.16
+pycparser==3.0
+tensordict==0.6.2
+airportsdata==20260315
+annotated-types==0.8.0
+nvidia-cudnn-cu12==9.1.0.70
+xformers==0.0.29.post2
+transformers==4.57.6
+pydantic-extra-types==2.11.1
+charset-normalizer==3.5.1
+python-dotenv==1.2.3
+nvidia-cuda-cupti-cu12==12.4.127
+mlflow-skinny==3.16.0
+packaging==26.3
+partial-json-parser==0.2.1.1.post7
+itsdangerous==2.2.0
+pydantic_core==2.46.5
+Jinja2==3.1.6
+fastapi-cloud-cli==0.26.0
+urllib3==2.7.0
+filelock==3.32.6
+wrapt==2.4.1
+opentelemetry-semantic-conventions-ai==0.4.13
+zss==1.2.0
+platformdirs==4.11.8
+importlib_metadata==8.0.0
+pytz==2026.3.post1
+google-auth==2.58.0
+nvidia-nvtx-cu12==12.4.127
+proto-plus==1.28.2
+shellingham==1.5.4
+attrs==26.1.0
+conda-pack==0.9.2
+vllm==0.8.5
+email-validator==2.3.0
+graphene==3.4.3
+pycosat==0.6.6
+magiccube==0.3.0
+torch==2.6.0
+pip==26.2.1
+rignore==0.8.1
+tzdata==2026.3
+wheel==0.47.0
+datasets==5.0.1
+joblib==1.6.0
+wandb==0.28.0
+mdurl==0.1.2
+googleapis-common-protos==1.75.0
+six==1.17.0
+graphql-core==3.2.12
+smart_open==8.0.1
+matplotlib==3.10.9
+opentelemetry-semantic-conventions==0.47b0
+idna==3.19
+omegaconf==2.3.1
+aiohappyeyeballs==2.7.1
+pyzmq==27.2.0
+sympy==1.13.1
+fsspec==2026.6.0
+fastapi==0.141.1
+fonttools==4.65.0
+numba==0.61.2
+annotated-doc==0.0.5
+databricks-sdk==0.138.0
+dill==0.4.1
+xxhash==4.0.1
+pyasn1==0.6.4
+tokenizers==0.22.2
+codetiming==1.4.0
+uvicorn==0.52.4
+latex2sympy2_extended==1.11.0
+opentelemetry-sdk==1.26.0
+docker==7.2.0
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+cffi==2.1.1
+starlette==1.6.0
+opentelemetry-exporter-otlp-proto-common==1.26.0
+typer==0.27.2
+torchvision==0.21.0
+wcwidth==0.8.3
+reasoning_gym==0.1.25
+blake3==1.0.9
+anyio==4.15.1
+cryptography==50.0.1
+greenlet==3.5.5
+nvidia-nvjitlink-cu12==12.4.127
+outlines==0.1.11
+flask-cors==6.0.5
+dnspython==2.8.0
+pillow==12.3.0
+contourpy==1.3.2
+opentelemetry-proto==1.26.0
+tomli==2.4.1
+networkx==3.4.2
+httpcore2==2.12.0
+opentelemetry-api==1.26.0
+nvidia-cufft-cu12==11.2.1.3
+nvidia-nccl-cu12==2.21.5
+Mako==1.4.1
+python-multipart==0.0.32
+cycler==0.12.1
+cellpylib==2.4.0
+gguf==0.19.0
+opencensus-context==0.1.3
+rich-toolkit==0.20.5
+sentencepiece==0.2.2
+fastapi-cli==0.0.32
+prometheus_client==0.26.0
+virtualenv==21.7.9
+certifi==2026.7.22
+Deprecated==1.3.1
+smmap==5.0.3
+watchfiles==1.2.0
+py-cpuinfo==9.0.0
+graphql-relay==3.2.0
+lark==1.2.2
+prettytable==3.18.0
+nvidia-curand-cu12==10.3.5.147
+bfi==1.0.4
+grpcio==1.83.1
+mlflow==3.16.0
+drawsvg==2.4.2
+py-spy==0.4.2
+tabulate==0.9.0
+aiohttp==3.14.3
+peft==0.20.0
+opentelemetry-exporter-otlp==1.26.0
+opencensus==0.11.4
+regex==2026.9.10
+Werkzeug==3.1.8
+typing_extensions==4.16.0
+truststore==0.10.4
+mlflow-tracing==3.16.0
+msgspec==0.21.1
+multidict==6.8.0
+nest-asyncio==1.6.0
+sentry-sdk==2.69.1
+async-timeout==5.0.1
+threadpoolctl==3.6.0
+h11==0.16.0
+httpx2==2.12.0
+multiprocess==0.70.19
+typing-inspection==0.4.4
+compressed-tensors==0.9.3
+httpcore==1.0.9
+llguidance==0.7.30
+more-itertools==10.3.0
+platformdirs==4.2.2
+jaraco.context==5.3.0
+inflect==7.3.1
+tomli==2.0.1
+autocommand==2.2.2
+zipp==3.19.2
+typing_extensions==4.12.2
+packaging==24.2
+importlib_metadata==8.0.0
+backports.tarfile==1.2.0
+jaraco.collections==5.1.0
+jaraco.functools==4.0.1
+wheel==0.45.1
+jaraco.text==3.12.1
+typeguard==4.3.0

training/llama3b_sft_rp_hard_lam1/wandb/run-20260923_014857-79qp1zp0/files/wandb-metadata.json

@@ -0,0 +1,125 @@
+{
+  "os": "Linux-6.8.0-1046-azure-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.21",
+  "startedAt": "2026-09-23T01:48:57.607753Z",
+  "args": [
+    "--dataset",
+    "llama3b_code_sft_data",
+    "--replay_strategy",
+    "hard",
+    "--replay_lambda",
+    "1.0",
+    "--replay_cooldown_step",
+    "3",
+    "--replay_ce_tau",
+    "1.0",
+    "--replay_ce_init",
+    "insert",
+    "--replay_group_size",
+    "8",
+    "--model_path",
+    "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+    "--output_dir",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922",
+    "--ordered",
+    "--system_prompt_mode",
+    "from_record",
+    "--lr",
+    "1e-5",
+    "--batch_size",
+    "1",
+    "--effective_batch_size",
+    "256",
+    "--max_seq_length",
+    "4096",
+    "--ddp_timeout_seconds",
+    "3600",
+    "--schedule",
+    "constant",
+    "--warmup_ratio",
+    "0",
+    "--num_epochs",
+    "1",
+    "--seed",
+    "42",
+    "--log_cuda_memory",
+    "--num_checkpoints",
+    "10",
+    "--save_optimizer_every_n_ckpts",
+    "5",
+    "--prune_old_optimizer",
+    "--save_optimizer_first_ckpt",
+    "--metrics_log_every",
+    "10",
+    "--precision",
+    "mixed_bf16",
+    "--grad_clip",
+    "0",
+    "--snapshot_device",
+    "cpu",
+    "--snapshot_dtype",
+    "float32",
+    "--wandb_project",
+    "rl_forgetting_sft",
+    "--wandb_entity",
+    "jingyanshen-new-york-university",
+    "--wandb_group",
+    "code_sft_20260922_llama3b",
+    "--wandb_run_name",
+    "llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922_segment_g1",
+    "--wandb_run_id_file",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt"
+  ],
+  "program": "/scratch/amlt_code/sft/run_sft.py",
+  "codePath": "sft/run_sft.py",
+  "codePathLocal": "sft/run_sft.py",
+  "email": "js15262@nyu.edu",
+  "root": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08096-3568aea7-abb6/llama3b_code_ordered_lr1e-5_bs256_seed42_rphard_lam1.0_gcr_20260922",
+  "host": "node-0",
+  "executable": "/tmp/rlf-code-runtime/bin/python",
+  "cpu_count": 96,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "30601604947968",
+      "used": "578502160384"
+    }
+  },
+  "memory": {
+    "total": "2008079052800"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-3b56a237-fc29-56a1-d432-428ca67ad7e4"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-6056d837-d995-d622-77fc-64519b1b237a"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-5099ed2a-4aa9-e0a3-4b19-55d9f18ca1f6"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-947282df-cba4-b58d-0319-dc7d31bb3cc9"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "4gm1jbz4pougcac6c2xr1sw2b6quaw96"
+}

training/llama3b_sft_rp_hard_lam1/wandb/run-20260923_014857-79qp1zp0/files/wandb-summary.json

@@ -0,0 +1 @@
+{"num_supervised_tokens":4748,"replay_draw_min_q":0.00015422578655151142,"cuda_current_allocated_gib":47.93620491027832,"parameter_l2_from_init":6.079890918161503,"_runtime":1808.834819278,"learning_rate":1e-05,"relative_parameter_l2_from_init":0.005470469846062616,"replay_draw_max_q":0.00107958050586058,"cuda_peak_allocated_gib":71.87305641174316,"batch_loss":0.21864656124753593,"replay_lambda":1,"replay_draw_n_groups":2705,"replay_draw_mean_pass_rate":0.625,"_timestamp":1.7901298124946125e+09,"replay_num_supervised_tokens":7824,"replay_loss":0.37019532127472515,"batch_id":4105,"relative_update_norm":0.0005490049924212632,"_wandb":{"runtime":1808},"parameter_l2_norm":1111.388887786505,"gradient_norm_pre_clip":3.6809591849601584,"_step":65,"epoch":0,"cuda_peak_reserved_gib":73.154296875,"gradient_norm_post_clip":3.6809591849601584,"global_step":65,"buffer_rows":4106,"token_entropy":0.20604007467062613,"update_norm":0.6101580479163062,"buffer_prompts":270,"total_loss":0.5888418825222611,"cuda_current_reserved_gib":73.154296875,"max_sequence_length":704}

training/llama3b_sft_rp_hard_lam1/wandb_run_id_generation_1.txt

@@ -0,0 +1 @@
+79qp1zp0

training/llama3b_sft_rp_hard_lam1/wandb_segments/generation_1.json

@@ -0,0 +1,8 @@
+{
+  "attempt_id": "yellow_arm_6lw7gf1fvy|node-0|20be5e94bbbf4028b32b403c38504cce",
+  "created_at": "2026-09-23T01:47:31.264440+00:00",
+  "generation": 1,
+  "resume_step": 0,
+  "schema_version": 1,
+  "wandb_run_id": "79qp1zp0"
+}

training/llama3b_sft_rp_uniform_lam1/metrics/train_metrics.jsonl

@@ -0,0 +1,65 @@
+{"global_step": 1, "epoch": 0, "batch_id": 63, "num_supervised_tokens": 97674, "learning_rate": 1e-05, "batch_loss": 1.3036121778480152, "token_entropy": 1.3033737616757028, "gradient_norm_pre_clip": 1.6957774339607559, "gradient_norm_post_clip": 1.6957774339607559, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.13671875, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.13671875, "max_sequence_length": 3030.0, "replay_loss": NaN, "replay_num_supervised_tokens": 0, "total_loss": 1.3036121778480152, "replay_lambda": 1.0, "buffer_rows": 64, "buffer_prompts": 54}
+{"global_step": 2, "epoch": 0, "batch_id": 127, "num_supervised_tokens": 92723, "learning_rate": 1e-05, "batch_loss": 0.6895444210004315, "token_entropy": 0.6529375377465082, "gradient_norm_pre_clip": 3.540591634657045, "gradient_norm_post_clip": 3.540591634657045, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.17578125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.17578125, "max_sequence_length": 2607.0, "replay_loss": 1.1441610864242586, "replay_num_supervised_tokens": 95261, "total_loss": 1.83370550742469, "replay_lambda": 1.0, "buffer_rows": 128, "buffer_prompts": 107, "replay_draw_n_groups": 54, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.018518518518518517, "replay_draw_max_q": 0.018518518518518517}
+{"global_step": 3, "epoch": 0, "batch_id": 191, "num_supervised_tokens": 92641, "learning_rate": 1e-05, "batch_loss": 0.7316602675151905, "token_entropy": 0.7672286962433656, "gradient_norm_pre_clip": 9.878170157671589, "gradient_norm_post_clip": 9.878170157671589, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.177734375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.177734375, "max_sequence_length": 1418.0, "replay_loss": 0.8544162876148503, "replay_num_supervised_tokens": 93617, "total_loss": 1.5860765551300409, "replay_lambda": 1.0, "buffer_rows": 192, "buffer_prompts": 127, "replay_draw_n_groups": 107, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.009345794392523364, "replay_draw_max_q": 0.009345794392523364}
+{"global_step": 4, "epoch": 0, "batch_id": 255, "num_supervised_tokens": 87718, "learning_rate": 1e-05, "batch_loss": 0.9076772054281532, "token_entropy": 0.9265057359756883, "gradient_norm_pre_clip": 2.547639942037403, "gradient_norm_post_clip": 2.547639942037403, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.197265625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.197265625, "max_sequence_length": 2607.0, "replay_loss": 1.3614321035182262, "replay_num_supervised_tokens": 100099, "total_loss": 2.2691093089463794, "replay_lambda": 1.0, "buffer_rows": 256, "buffer_prompts": 148, "replay_draw_n_groups": 127, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.007874015748031496, "replay_draw_max_q": 0.007874015748031496}
+{"global_step": 5, "epoch": 0, "batch_id": 319, "num_supervised_tokens": 82516, "learning_rate": 1e-05, "batch_loss": 0.8512886792830908, "token_entropy": 0.7945125968765199, "gradient_norm_pre_clip": 2.9933582617024266, "gradient_norm_post_clip": 2.9933582617024266, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.197265625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.197265625, "max_sequence_length": 2017.0, "replay_loss": 0.6085882397230853, "replay_num_supervised_tokens": 88183, "total_loss": 1.459876919006176, "replay_lambda": 1.0, "buffer_rows": 320, "buffer_prompts": 168, "replay_draw_n_groups": 148, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.006756756756756757, "replay_draw_max_q": 0.006756756756756757}
+{"global_step": 6, "epoch": 0, "batch_id": 383, "num_supervised_tokens": 78503, "learning_rate": 1e-05, "batch_loss": 1.15251624207754, "token_entropy": 1.113781188519343, "gradient_norm_pre_clip": 2.280845563457089, "gradient_norm_post_clip": 2.280845563457089, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.197265625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.197265625, "max_sequence_length": 2750.0, "replay_loss": 0.5103409759510402, "replay_num_supervised_tokens": 85493, "total_loss": 1.6628572180285803, "replay_lambda": 1.0, "buffer_rows": 384, "buffer_prompts": 182, "replay_draw_n_groups": 168, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.005952380952380952, "replay_draw_max_q": 0.005952380952380952}
+{"global_step": 7, "epoch": 0, "batch_id": 447, "num_supervised_tokens": 75606, "learning_rate": 1e-05, "batch_loss": 1.1952957567544449, "token_entropy": 1.1901025653021036, "gradient_norm_pre_clip": 1.4365863642685317, "gradient_norm_post_clip": 1.4365863642685317, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2916.0, "replay_loss": 1.0246360124558977, "replay_num_supervised_tokens": 93906, "total_loss": 2.2199317692103424, "replay_lambda": 1.0, "buffer_rows": 448, "buffer_prompts": 190, "replay_draw_n_groups": 182, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.005494505494505495, "replay_draw_max_q": 0.005494505494505495}
+{"global_step": 8, "epoch": 0, "batch_id": 511, "num_supervised_tokens": 65460, "learning_rate": 1e-05, "batch_loss": 0.6346750233817494, "token_entropy": 0.6440822658428096, "gradient_norm_pre_clip": 1.8045837349431606, "gradient_norm_post_clip": 1.8045837349431606, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 1174.0, "replay_loss": 0.683712577916016, "replay_num_supervised_tokens": 88733, "total_loss": 1.3183876012977653, "replay_lambda": 1.0, "buffer_rows": 512, "buffer_prompts": 197, "replay_draw_n_groups": 190, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.005263157894736842, "replay_draw_max_q": 0.005263157894736842}
+{"global_step": 9, "epoch": 0, "batch_id": 575, "num_supervised_tokens": 60380, "learning_rate": 1e-05, "batch_loss": 0.6583545487876284, "token_entropy": 0.6754384741162417, "gradient_norm_pre_clip": 1.8194924077866712, "gradient_norm_post_clip": 1.8194924077866712, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2871.0, "replay_loss": 1.4380236079277278, "replay_num_supervised_tokens": 90611, "total_loss": 2.0963781567153563, "replay_lambda": 1.0, "buffer_rows": 576, "buffer_prompts": 200, "replay_draw_n_groups": 197, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.005076142131979695, "replay_draw_max_q": 0.005076142131979695}
+{"global_step": 10, "epoch": 0, "batch_id": 639, "num_supervised_tokens": 60964, "learning_rate": 1e-05, "batch_loss": 0.5184750687271077, "token_entropy": 0.5329639531677378, "gradient_norm_pre_clip": 1.540062636261316, "gradient_norm_post_clip": 1.540062636261316, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2066.0, "replay_loss": 0.7302999246484306, "replay_num_supervised_tokens": 82235, "total_loss": 1.2487749933755383, "replay_lambda": 1.0, "buffer_rows": 640, "buffer_prompts": 204, "replay_draw_n_groups": 200, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.005, "replay_draw_max_q": 0.005, "parameter_l2_norm": 1111.3988448294374, "parameter_l2_from_init": 2.2097449452155344, "relative_parameter_l2_from_init": 0.001988249995436803, "update_norm": 2.2097449452155344, "relative_update_norm": 0.001988255571342308}
+{"global_step": 11, "epoch": 0, "batch_id": 703, "num_supervised_tokens": 54290, "learning_rate": 1e-05, "batch_loss": 0.47942369586718986, "token_entropy": 0.4709315293839589, "gradient_norm_pre_clip": 1.5116222738063105, "gradient_norm_post_clip": 1.5116222738063105, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2607.0, "replay_loss": 0.786583757802248, "replay_num_supervised_tokens": 80319, "total_loss": 1.266007453669438, "replay_lambda": 1.0, "buffer_rows": 704, "buffer_prompts": 206, "replay_draw_n_groups": 204, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.004901960784313725, "replay_draw_max_q": 0.004901960784313725}
+{"global_step": 12, "epoch": 0, "batch_id": 767, "num_supervised_tokens": 54744, "learning_rate": 1e-05, "batch_loss": 0.4299656982755066, "token_entropy": 0.4136141989291385, "gradient_norm_pre_clip": 1.6916109242767006, "gradient_norm_post_clip": 1.6916109242767006, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2066.0, "replay_loss": 0.8449903865543206, "replay_num_supervised_tokens": 80230, "total_loss": 1.2749560848298271, "replay_lambda": 1.0, "buffer_rows": 768, "buffer_prompts": 209, "replay_draw_n_groups": 206, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0048543689320388345, "replay_draw_max_q": 0.0048543689320388345}
+{"global_step": 13, "epoch": 0, "batch_id": 831, "num_supervised_tokens": 48778, "learning_rate": 1e-05, "batch_loss": 0.4346153885084873, "token_entropy": 0.42655622536339655, "gradient_norm_pre_clip": 1.575514381608115, "gradient_norm_post_clip": 1.575514381608115, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 978.0, "replay_loss": 0.6293086663301118, "replay_num_supervised_tokens": 73135, "total_loss": 1.063924054838599, "replay_lambda": 1.0, "buffer_rows": 832, "buffer_prompts": 212, "replay_draw_n_groups": 209, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.004784688995215311, "replay_draw_max_q": 0.004784688995215311}
+{"global_step": 14, "epoch": 0, "batch_id": 895, "num_supervised_tokens": 48298, "learning_rate": 1e-05, "batch_loss": 0.40814534834001004, "token_entropy": 0.41042929901346725, "gradient_norm_pre_clip": 1.4826339457795592, "gradient_norm_post_clip": 1.4826339457795592, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.861328125, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.861328125, "max_sequence_length": 2017.0, "replay_loss": 0.8025844011199804, "replay_num_supervised_tokens": 71007, "total_loss": 1.2107297494599905, "replay_lambda": 1.0, "buffer_rows": 896, "buffer_prompts": 214, "replay_draw_n_groups": 212, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.0047169811320754715, "replay_draw_max_q": 0.0047169811320754715}
+{"global_step": 15, "epoch": 0, "batch_id": 959, "num_supervised_tokens": 44942, "learning_rate": 1e-05, "batch_loss": 0.3537931666141223, "token_entropy": 0.3735473738149141, "gradient_norm_pre_clip": 1.5249921526070824, "gradient_norm_post_clip": 1.5249921526070824, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.880859375, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.880859375, "max_sequence_length": 2916.0, "replay_loss": 0.9349286119149444, "replay_num_supervised_tokens": 78363, "total_loss": 1.2887217785290668, "replay_lambda": 1.0, "buffer_rows": 960, "buffer_prompts": 216, "replay_draw_n_groups": 214, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.004672897196261682, "replay_draw_max_q": 0.004672897196261682}
+{"global_step": 16, "epoch": 0, "batch_id": 1023, "num_supervised_tokens": 40349, "learning_rate": 1e-05, "batch_loss": 0.34002356216938906, "token_entropy": 0.3635925238228615, "gradient_norm_pre_clip": 1.5382400076564988, "gradient_norm_post_clip": 1.5382400076564988, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 3030.0, "replay_loss": 0.7270265537432098, "replay_num_supervised_tokens": 73673, "total_loss": 1.067050115912599, "replay_lambda": 1.0, "buffer_rows": 1024, "buffer_prompts": 218, "replay_draw_n_groups": 216, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.004629629629629629, "replay_draw_max_q": 0.004629629629629629}
+{"global_step": 17, "epoch": 0, "batch_id": 1087, "num_supervised_tokens": 41901, "learning_rate": 1e-05, "batch_loss": 0.32481935016398544, "token_entropy": 0.3453251615247392, "gradient_norm_pre_clip": 1.4214905601838226, "gradient_norm_post_clip": 1.4214905601838226, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 906.0, "replay_loss": 0.49989604864671855, "replay_num_supervised_tokens": 67884, "total_loss": 0.824715398810704, "replay_lambda": 1.0, "buffer_rows": 1088, "buffer_prompts": 223, "replay_draw_n_groups": 218, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.0045871559633027525, "replay_draw_max_q": 0.0045871559633027525}
+{"global_step": 18, "epoch": 0, "batch_id": 1151, "num_supervised_tokens": 37655, "learning_rate": 1e-05, "batch_loss": 0.3490277381251017, "token_entropy": 0.35564055277553785, "gradient_norm_pre_clip": 1.4879768232265382, "gradient_norm_post_clip": 1.4879768232265382, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2871.0, "replay_loss": 1.0674614737906967, "replay_num_supervised_tokens": 70639, "total_loss": 1.4164892119157984, "replay_lambda": 1.0, "buffer_rows": 1152, "buffer_prompts": 225, "replay_draw_n_groups": 223, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.004484304932735426, "replay_draw_max_q": 0.004484304932735426}
+{"global_step": 19, "epoch": 0, "batch_id": 1215, "num_supervised_tokens": 38084, "learning_rate": 1e-05, "batch_loss": 0.29155586941818984, "token_entropy": 0.2945609031110711, "gradient_norm_pre_clip": 1.5924911089171716, "gradient_norm_post_clip": 1.5924911089171716, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2607.0, "replay_loss": 1.3355377067965222, "replay_num_supervised_tokens": 73554, "total_loss": 1.627093576214712, "replay_lambda": 1.0, "buffer_rows": 1216, "buffer_prompts": 226, "replay_draw_n_groups": 225, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0044444444444444444, "replay_draw_max_q": 0.0044444444444444444}
+{"global_step": 20, "epoch": 0, "batch_id": 1279, "num_supervised_tokens": 36949, "learning_rate": 1e-05, "batch_loss": 0.28134707928324115, "token_entropy": 0.2952204273790367, "gradient_norm_pre_clip": 1.6675393358184751, "gradient_norm_post_clip": 1.6675393358184751, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1055.0, "replay_loss": 0.4961305238834797, "replay_num_supervised_tokens": 62620, "total_loss": 0.7774776031667209, "replay_lambda": 1.0, "buffer_rows": 1280, "buffer_prompts": 229, "replay_draw_n_groups": 226, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.004424778761061947, "replay_draw_max_q": 0.004424778761061947, "parameter_l2_norm": 1111.3962573252895, "parameter_l2_from_init": 3.3213015242739683, "relative_parameter_l2_from_init": 0.0029883891146712696, "update_norm": 1.458740106135227, "relative_update_norm": 0.00131252925904741}
+{"global_step": 21, "epoch": 0, "batch_id": 1343, "num_supervised_tokens": 36603, "learning_rate": 1e-05, "batch_loss": 0.28480507578383546, "token_entropy": 0.29258309607068994, "gradient_norm_pre_clip": 1.5010851905662108, "gradient_norm_post_clip": 1.5010851905662108, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 756.0, "replay_loss": 0.4437952805712276, "replay_num_supervised_tokens": 59293, "total_loss": 0.728600356355063, "replay_lambda": 1.0, "buffer_rows": 1344, "buffer_prompts": 232, "replay_draw_n_groups": 229, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.004366812227074236, "replay_draw_max_q": 0.004366812227074236}
+{"global_step": 22, "epoch": 0, "batch_id": 1407, "num_supervised_tokens": 36508, "learning_rate": 1e-05, "batch_loss": 0.26221181324111603, "token_entropy": 0.2702491280020844, "gradient_norm_pre_clip": 1.5644338768045443, "gradient_norm_post_clip": 1.5644338768045443, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 807.0, "replay_loss": 0.38028429258766244, "replay_num_supervised_tokens": 63898, "total_loss": 0.6424961058287785, "replay_lambda": 1.0, "buffer_rows": 1408, "buffer_prompts": 233, "replay_draw_n_groups": 232, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.004310344827586207, "replay_draw_max_q": 0.004310344827586207}
+{"global_step": 23, "epoch": 0, "batch_id": 1471, "num_supervised_tokens": 34482, "learning_rate": 1e-05, "batch_loss": 0.2816754414333568, "token_entropy": 0.2847787183769673, "gradient_norm_pre_clip": 1.5856351993858462, "gradient_norm_post_clip": 1.5856351993858462, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1174.0, "replay_loss": 0.45751057551461577, "replay_num_supervised_tokens": 59385, "total_loss": 0.7391860169479726, "replay_lambda": 1.0, "buffer_rows": 1472, "buffer_prompts": 233, "replay_draw_n_groups": 233, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.004291845493562232, "replay_draw_max_q": 0.004291845493562232}
+{"global_step": 24, "epoch": 0, "batch_id": 1535, "num_supervised_tokens": 34610, "learning_rate": 1e-05, "batch_loss": 0.26767156677752757, "token_entropy": 0.2736551801168447, "gradient_norm_pre_clip": 1.6221590250345421, "gradient_norm_post_clip": 1.6221590250345421, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2916.0, "replay_loss": 0.9402387612928049, "replay_num_supervised_tokens": 64307, "total_loss": 1.2079103280703325, "replay_lambda": 1.0, "buffer_rows": 1536, "buffer_prompts": 235, "replay_draw_n_groups": 233, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.004291845493562232, "replay_draw_max_q": 0.004291845493562232}
+{"global_step": 25, "epoch": 0, "batch_id": 1599, "num_supervised_tokens": 36993, "learning_rate": 1e-05, "batch_loss": 0.24564187086467482, "token_entropy": 0.2433417185706382, "gradient_norm_pre_clip": 1.428860665776448, "gradient_norm_post_clip": 1.428860665776448, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 856.0, "replay_loss": 0.3733546232334234, "replay_num_supervised_tokens": 60446, "total_loss": 0.6189964940980982, "replay_lambda": 1.0, "buffer_rows": 1600, "buffer_prompts": 235, "replay_draw_n_groups": 235, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.00425531914893617, "replay_draw_max_q": 0.00425531914893617}
+{"global_step": 26, "epoch": 0, "batch_id": 1663, "num_supervised_tokens": 35574, "learning_rate": 1e-05, "batch_loss": 0.2481667913792641, "token_entropy": 0.24168414525910117, "gradient_norm_pre_clip": 1.7560779546602345, "gradient_norm_post_clip": 1.7560779546602345, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 994.0, "replay_loss": 0.32615480500065763, "replay_num_supervised_tokens": 54792, "total_loss": 0.5743215963799218, "replay_lambda": 1.0, "buffer_rows": 1664, "buffer_prompts": 236, "replay_draw_n_groups": 235, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.00425531914893617, "replay_draw_max_q": 0.00425531914893617}
+{"global_step": 27, "epoch": 0, "batch_id": 1727, "num_supervised_tokens": 31824, "learning_rate": 1e-05, "batch_loss": 0.24909194833385984, "token_entropy": 0.24446954638470111, "gradient_norm_pre_clip": 1.6105209681669157, "gradient_norm_post_clip": 1.6105209681669157, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 763.0, "replay_loss": 0.3526100908744859, "replay_num_supervised_tokens": 53983, "total_loss": 0.6017020392083458, "replay_lambda": 1.0, "buffer_rows": 1728, "buffer_prompts": 237, "replay_draw_n_groups": 236, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.00423728813559322, "replay_draw_max_q": 0.00423728813559322}
+{"global_step": 28, "epoch": 0, "batch_id": 1791, "num_supervised_tokens": 33894, "learning_rate": 1e-05, "batch_loss": 0.238084558714512, "token_entropy": 0.24122003510953993, "gradient_norm_pre_clip": 1.5398197877468618, "gradient_norm_post_clip": 1.5398197877468618, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2017.0, "replay_loss": 0.6478854801327002, "replay_num_supervised_tokens": 56233, "total_loss": 0.8859700388472123, "replay_lambda": 1.0, "buffer_rows": 1792, "buffer_prompts": 242, "replay_draw_n_groups": 237, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.004219409282700422, "replay_draw_max_q": 0.004219409282700422}
+{"global_step": 29, "epoch": 0, "batch_id": 1855, "num_supervised_tokens": 32475, "learning_rate": 1e-05, "batch_loss": 0.2299095722927416, "token_entropy": 0.24368140851285844, "gradient_norm_pre_clip": 1.758625904840092, "gradient_norm_post_clip": 1.758625904840092, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1101.0, "replay_loss": 0.30912610050636213, "replay_num_supervised_tokens": 57164, "total_loss": 0.5390356727991037, "replay_lambda": 1.0, "buffer_rows": 1856, "buffer_prompts": 243, "replay_draw_n_groups": 242, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.004132231404958678, "replay_draw_max_q": 0.004132231404958678}
+{"global_step": 30, "epoch": 0, "batch_id": 1919, "num_supervised_tokens": 30711, "learning_rate": 1e-05, "batch_loss": 0.23743555666482466, "token_entropy": 0.24558738495968108, "gradient_norm_pre_clip": 1.5688938165012787, "gradient_norm_post_clip": 1.5688938165012787, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 819.0, "replay_loss": 0.3038087882669592, "replay_num_supervised_tokens": 52044, "total_loss": 0.5412443449317839, "replay_lambda": 1.0, "buffer_rows": 1920, "buffer_prompts": 243, "replay_draw_n_groups": 243, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00411522633744856, "replay_draw_max_q": 0.00411522633744856, "parameter_l2_norm": 1111.3937981783106, "parameter_l2_from_init": 4.07943062796341, "relative_parameter_l2_from_init": 0.0036705267478921097, "update_norm": 1.1906159077313947, "relative_update_norm": 0.0010712817632084481}
+{"global_step": 31, "epoch": 0, "batch_id": 1983, "num_supervised_tokens": 31193, "learning_rate": 1e-05, "batch_loss": 0.204857830079419, "token_entropy": 0.21163019658403803, "gradient_norm_pre_clip": 1.569639508459932, "gradient_norm_post_clip": 1.569639508459932, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1139.0, "replay_loss": 0.35248480973915247, "replay_num_supervised_tokens": 52691, "total_loss": 0.5573426398185715, "replay_lambda": 1.0, "buffer_rows": 1984, "buffer_prompts": 245, "replay_draw_n_groups": 243, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00411522633744856, "replay_draw_max_q": 0.00411522633744856}
+{"global_step": 32, "epoch": 0, "batch_id": 2047, "num_supervised_tokens": 29719, "learning_rate": 1e-05, "batch_loss": 0.2037536173458736, "token_entropy": 0.2060822576696561, "gradient_norm_pre_clip": 1.6738631148138192, "gradient_norm_post_clip": 1.6738631148138192, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2916.0, "replay_loss": 0.740165176107627, "replay_num_supervised_tokens": 52610, "total_loss": 0.9439187934535006, "replay_lambda": 1.0, "buffer_rows": 2048, "buffer_prompts": 246, "replay_draw_n_groups": 245, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.004081632653061225, "replay_draw_max_q": 0.004081632653061225}
+{"global_step": 33, "epoch": 0, "batch_id": 2111, "num_supervised_tokens": 30714, "learning_rate": 1e-05, "batch_loss": 0.19046332963146237, "token_entropy": 0.18884416044771404, "gradient_norm_pre_clip": 1.6027953973950744, "gradient_norm_post_clip": 1.6027953973950744, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 888.0, "replay_loss": 0.3440679319420936, "replay_num_supervised_tokens": 53854, "total_loss": 0.534531261573556, "replay_lambda": 1.0, "buffer_rows": 2112, "buffer_prompts": 247, "replay_draw_n_groups": 246, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0040650406504065045, "replay_draw_max_q": 0.0040650406504065045}
+{"global_step": 34, "epoch": 0, "batch_id": 2175, "num_supervised_tokens": 29442, "learning_rate": 1e-05, "batch_loss": 0.16821427838926822, "token_entropy": 0.17177076017748896, "gradient_norm_pre_clip": 1.6729973270149663, "gradient_norm_post_clip": 1.6729973270149663, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 742.0, "replay_loss": 0.30876327465194603, "replay_num_supervised_tokens": 51504, "total_loss": 0.47697755304121425, "replay_lambda": 1.0, "buffer_rows": 2176, "buffer_prompts": 248, "replay_draw_n_groups": 247, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.004048582995951417, "replay_draw_max_q": 0.004048582995951417}
+{"global_step": 35, "epoch": 0, "batch_id": 2239, "num_supervised_tokens": 34180, "learning_rate": 1e-05, "batch_loss": 0.15468527402983834, "token_entropy": 0.16117030847414396, "gradient_norm_pre_clip": 1.4756306533075874, "gradient_norm_post_clip": 1.4756306533075874, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1052.0, "replay_loss": 0.29745769231676716, "replay_num_supervised_tokens": 50533, "total_loss": 0.4521429663466055, "replay_lambda": 1.0, "buffer_rows": 2240, "buffer_prompts": 250, "replay_draw_n_groups": 248, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.004032258064516129, "replay_draw_max_q": 0.004032258064516129}
+{"global_step": 36, "epoch": 0, "batch_id": 2303, "num_supervised_tokens": 30222, "learning_rate": 1e-05, "batch_loss": 0.16024104610046536, "token_entropy": 0.172679491436274, "gradient_norm_pre_clip": 1.6101851627303927, "gradient_norm_post_clip": 1.6101851627303927, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1436.0, "replay_loss": 0.4795717513370483, "replay_num_supervised_tokens": 54210, "total_loss": 0.6398127974375136, "replay_lambda": 1.0, "buffer_rows": 2304, "buffer_prompts": 254, "replay_draw_n_groups": 250, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.004, "replay_draw_max_q": 0.004}
+{"global_step": 37, "epoch": 0, "batch_id": 2367, "num_supervised_tokens": 31163, "learning_rate": 1e-05, "batch_loss": 0.15402024921803362, "token_entropy": 0.16548863977621314, "gradient_norm_pre_clip": 1.5196536698999377, "gradient_norm_post_clip": 1.5196536698999377, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 875.0, "replay_loss": 0.30711647872672565, "replay_num_supervised_tokens": 46771, "total_loss": 0.46113672794475924, "replay_lambda": 1.0, "buffer_rows": 2368, "buffer_prompts": 256, "replay_draw_n_groups": 254, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.003937007874015748, "replay_draw_max_q": 0.003937007874015748}
+{"global_step": 38, "epoch": 0, "batch_id": 2431, "num_supervised_tokens": 28931, "learning_rate": 1e-05, "batch_loss": 0.15637302376727266, "token_entropy": 0.1685150077985698, "gradient_norm_pre_clip": 1.5163812844037177, "gradient_norm_post_clip": 1.5163812844037177, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2607.0, "replay_loss": 0.6563164146536594, "replay_num_supervised_tokens": 51045, "total_loss": 0.8126894384209321, "replay_lambda": 1.0, "buffer_rows": 2432, "buffer_prompts": 256, "replay_draw_n_groups": 256, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.00390625, "replay_draw_max_q": 0.00390625}
+{"global_step": 39, "epoch": 0, "batch_id": 2495, "num_supervised_tokens": 29457, "learning_rate": 1e-05, "batch_loss": 0.1352953035776459, "token_entropy": 0.15180765100112575, "gradient_norm_pre_clip": 1.5709866485558597, "gradient_norm_post_clip": 1.5709866485558597, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2066.0, "replay_loss": 0.584309434609684, "replay_num_supervised_tokens": 48569, "total_loss": 0.7196047381873298, "replay_lambda": 1.0, "buffer_rows": 2496, "buffer_prompts": 257, "replay_draw_n_groups": 256, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.00390625, "replay_draw_max_q": 0.00390625}
+{"global_step": 40, "epoch": 0, "batch_id": 2559, "num_supervised_tokens": 30001, "learning_rate": 1e-05, "batch_loss": 0.13093710974923486, "token_entropy": 0.13992764146275283, "gradient_norm_pre_clip": 1.5095805712109989, "gradient_norm_post_clip": 1.5095805712109989, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 847.0, "replay_loss": 0.26858452771371144, "replay_num_supervised_tokens": 46728, "total_loss": 0.39952163746294633, "replay_lambda": 1.0, "buffer_rows": 2560, "buffer_prompts": 258, "replay_draw_n_groups": 257, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0038910505836575876, "replay_draw_max_q": 0.0038910505836575876, "parameter_l2_norm": 1111.3910580235363, "parameter_l2_from_init": 4.63720686588351, "relative_parameter_l2_from_init": 0.004172393990489751, "update_norm": 1.0655145767746592, "relative_update_norm": 0.0009587215670688745}
+{"global_step": 41, "epoch": 0, "batch_id": 2623, "num_supervised_tokens": 29105, "learning_rate": 1e-05, "batch_loss": 0.13833533320988736, "token_entropy": 0.13461351730511822, "gradient_norm_pre_clip": 1.717033652096747, "gradient_norm_post_clip": 1.717033652096747, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1418.0, "replay_loss": 0.4451258556618931, "replay_num_supervised_tokens": 48642, "total_loss": 0.5834611888717804, "replay_lambda": 1.0, "buffer_rows": 2624, "buffer_prompts": 258, "replay_draw_n_groups": 258, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.003875968992248062, "replay_draw_max_q": 0.003875968992248062}
+{"global_step": 42, "epoch": 0, "batch_id": 2687, "num_supervised_tokens": 28971, "learning_rate": 1e-05, "batch_loss": 0.11518552592693823, "token_entropy": 0.12188549403190317, "gradient_norm_pre_clip": 1.6851202719841458, "gradient_norm_post_clip": 1.6851202719841458, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1174.0, "replay_loss": 0.35115123409186755, "replay_num_supervised_tokens": 46950, "total_loss": 0.46633676001880575, "replay_lambda": 1.0, "buffer_rows": 2688, "buffer_prompts": 259, "replay_draw_n_groups": 258, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.003875968992248062, "replay_draw_max_q": 0.003875968992248062}
+{"global_step": 43, "epoch": 0, "batch_id": 2751, "num_supervised_tokens": 28111, "learning_rate": 1e-05, "batch_loss": 0.11422682584844389, "token_entropy": 0.11615253124974596, "gradient_norm_pre_clip": 1.7586586463688385, "gradient_norm_post_clip": 1.7586586463688385, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 847.0, "replay_loss": 0.25865255502522766, "replay_num_supervised_tokens": 47580, "total_loss": 0.37287938087367156, "replay_lambda": 1.0, "buffer_rows": 2752, "buffer_prompts": 259, "replay_draw_n_groups": 259, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.003861003861003861, "replay_draw_max_q": 0.003861003861003861}
+{"global_step": 44, "epoch": 0, "batch_id": 2815, "num_supervised_tokens": 29323, "learning_rate": 1e-05, "batch_loss": 0.11607813052414473, "token_entropy": 0.11678331426703678, "gradient_norm_pre_clip": 1.7062801985770242, "gradient_norm_post_clip": 1.7062801985770242, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1052.0, "replay_loss": 0.3370619511026315, "replay_num_supervised_tokens": 47612, "total_loss": 0.4531400816267762, "replay_lambda": 1.0, "buffer_rows": 2816, "buffer_prompts": 259, "replay_draw_n_groups": 259, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.003861003861003861, "replay_draw_max_q": 0.003861003861003861}
+{"global_step": 45, "epoch": 0, "batch_id": 2879, "num_supervised_tokens": 26914, "learning_rate": 1e-05, "batch_loss": 0.1149242554176088, "token_entropy": 0.12108019460068775, "gradient_norm_pre_clip": 1.852389283421754, "gradient_norm_post_clip": 1.852389283421754, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 847.0, "replay_loss": 0.32395965222808715, "replay_num_supervised_tokens": 45601, "total_loss": 0.43888390764569596, "replay_lambda": 1.0, "buffer_rows": 2880, "buffer_prompts": 259, "replay_draw_n_groups": 259, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.003861003861003861, "replay_draw_max_q": 0.003861003861003861}
+{"global_step": 46, "epoch": 0, "batch_id": 2943, "num_supervised_tokens": 25515, "learning_rate": 1e-05, "batch_loss": 0.117235648351357, "token_entropy": 0.12229791443510989, "gradient_norm_pre_clip": 1.8178067670982252, "gradient_norm_post_clip": 1.8178067670982252, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 761.0, "replay_loss": 0.3598311856371376, "replay_num_supervised_tokens": 44907, "total_loss": 0.4770668339884946, "replay_lambda": 1.0, "buffer_rows": 2944, "buffer_prompts": 259, "replay_draw_n_groups": 259, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.003861003861003861, "replay_draw_max_q": 0.003861003861003861}
+{"global_step": 47, "epoch": 0, "batch_id": 3007, "num_supervised_tokens": 30018, "learning_rate": 1e-05, "batch_loss": 0.08926045056829611, "token_entropy": 0.09614966141088663, "gradient_norm_pre_clip": 1.69631148476822, "gradient_norm_post_clip": 1.69631148476822, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1418.0, "replay_loss": 0.6920934751447939, "replay_num_supervised_tokens": 47352, "total_loss": 0.78135392571309, "replay_lambda": 1.0, "buffer_rows": 3008, "buffer_prompts": 261, "replay_draw_n_groups": 259, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.003861003861003861, "replay_draw_max_q": 0.003861003861003861}
+{"global_step": 48, "epoch": 0, "batch_id": 3071, "num_supervised_tokens": 30651, "learning_rate": 1e-05, "batch_loss": 0.09425829349559751, "token_entropy": 0.0993326281782488, "gradient_norm_pre_clip": 1.6739769885486104, "gradient_norm_post_clip": 1.6739769885486104, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 937.0, "replay_loss": 0.2530552791283345, "replay_num_supervised_tokens": 46063, "total_loss": 0.347313572623932, "replay_lambda": 1.0, "buffer_rows": 3072, "buffer_prompts": 261, "replay_draw_n_groups": 261, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.0038314176245210726, "replay_draw_max_q": 0.0038314176245210726}
+{"global_step": 49, "epoch": 0, "batch_id": 3135, "num_supervised_tokens": 28413, "learning_rate": 1e-05, "batch_loss": 0.11603155343689131, "token_entropy": 0.11239985904794969, "gradient_norm_pre_clip": 1.897200445440945, "gradient_norm_post_clip": 1.897200445440945, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 836.0, "replay_loss": 0.3445408504830434, "replay_num_supervised_tokens": 46358, "total_loss": 0.4605724039199347, "replay_lambda": 1.0, "buffer_rows": 3136, "buffer_prompts": 261, "replay_draw_n_groups": 261, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.0038314176245210726, "replay_draw_max_q": 0.0038314176245210726}
+{"global_step": 50, "epoch": 0, "batch_id": 3199, "num_supervised_tokens": 27750, "learning_rate": 1e-05, "batch_loss": 0.0936179235107727, "token_entropy": 0.10553620381398245, "gradient_norm_pre_clip": 1.8445230746641492, "gradient_norm_post_clip": 1.8445230746641492, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1436.0, "replay_loss": 0.4754520292778855, "replay_num_supervised_tokens": 44956, "total_loss": 0.5690699527886582, "replay_lambda": 1.0, "buffer_rows": 3200, "buffer_prompts": 261, "replay_draw_n_groups": 261, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0038314176245210726, "replay_draw_max_q": 0.0038314176245210726, "parameter_l2_norm": 1111.3884222592876, "parameter_l2_from_init": 5.119684889459868, "relative_parameter_l2_from_init": 0.004606510574963026, "update_norm": 1.0358733716860027, "relative_update_norm": 0.0009320534125955946}
+{"global_step": 51, "epoch": 0, "batch_id": 3263, "num_supervised_tokens": 24101, "learning_rate": 1e-05, "batch_loss": 0.09907713240940487, "token_entropy": 0.1007213764262691, "gradient_norm_pre_clip": 1.7653153063906333, "gradient_norm_post_clip": 1.7653153063906333, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 2871.0, "replay_loss": 0.6843779447522665, "replay_num_supervised_tokens": 49320, "total_loss": 0.7834550771616714, "replay_lambda": 1.0, "buffer_rows": 3264, "buffer_prompts": 261, "replay_draw_n_groups": 261, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.0038314176245210726, "replay_draw_max_q": 0.0038314176245210726}
+{"global_step": 52, "epoch": 0, "batch_id": 3327, "num_supervised_tokens": 29272, "learning_rate": 1e-05, "batch_loss": 0.10032959560121467, "token_entropy": 0.0955203923738599, "gradient_norm_pre_clip": 1.8920617074210722, "gradient_norm_post_clip": 1.8920617074210722, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1174.0, "replay_loss": 0.3895407310410429, "replay_num_supervised_tokens": 42132, "total_loss": 0.4898703266422576, "replay_lambda": 1.0, "buffer_rows": 3328, "buffer_prompts": 262, "replay_draw_n_groups": 261, "replay_draw_mean_pass_rate": 0.75, "replay_draw_min_q": 0.0038314176245210726, "replay_draw_max_q": 0.0038314176245210726}
+{"global_step": 53, "epoch": 0, "batch_id": 3391, "num_supervised_tokens": 29206, "learning_rate": 1e-05, "batch_loss": 0.11773764465259565, "token_entropy": 0.09706318767839921, "gradient_norm_pre_clip": 3.062710140612982, "gradient_norm_post_clip": 3.062710140612982, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 715.0, "replay_loss": 0.23993975237867488, "replay_num_supervised_tokens": 39457, "total_loss": 0.35767739703127055, "replay_lambda": 1.0, "buffer_rows": 3392, "buffer_prompts": 263, "replay_draw_n_groups": 262, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.003816793893129771, "replay_draw_max_q": 0.003816793893129771}
+{"global_step": 54, "epoch": 0, "batch_id": 3455, "num_supervised_tokens": 28302, "learning_rate": 1e-05, "batch_loss": 0.09660147349765764, "token_entropy": 0.11621318888844522, "gradient_norm_pre_clip": 1.8016593949894826, "gradient_norm_post_clip": 1.8016593949894826, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 865.0, "replay_loss": 0.26432981019079727, "replay_num_supervised_tokens": 41708, "total_loss": 0.3609312836884549, "replay_lambda": 1.0, "buffer_rows": 3456, "buffer_prompts": 263, "replay_draw_n_groups": 263, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0038022813688212928, "replay_draw_max_q": 0.0038022813688212928}
+{"global_step": 55, "epoch": 0, "batch_id": 3519, "num_supervised_tokens": 24971, "learning_rate": 1e-05, "batch_loss": 0.10617042011685604, "token_entropy": 0.12695049858605026, "gradient_norm_pre_clip": 2.2563830561784224, "gradient_norm_post_clip": 2.2563830561784224, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 830.0, "replay_loss": 0.28219994950115024, "replay_num_supervised_tokens": 41547, "total_loss": 0.3883703696180063, "replay_lambda": 1.0, "buffer_rows": 3520, "buffer_prompts": 264, "replay_draw_n_groups": 263, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.0038022813688212928, "replay_draw_max_q": 0.0038022813688212928}
+{"global_step": 56, "epoch": 0, "batch_id": 3583, "num_supervised_tokens": 31094, "learning_rate": 1e-05, "batch_loss": 0.07802893048741899, "token_entropy": 0.10264966669915045, "gradient_norm_pre_clip": 1.7065814370731602, "gradient_norm_post_clip": 1.7065814370731602, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 847.0, "replay_loss": 0.34465571167296044, "replay_num_supervised_tokens": 42617, "total_loss": 0.42268464216037943, "replay_lambda": 1.0, "buffer_rows": 3584, "buffer_prompts": 264, "replay_draw_n_groups": 264, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.003787878787878788, "replay_draw_max_q": 0.003787878787878788}
+{"global_step": 57, "epoch": 0, "batch_id": 3647, "num_supervised_tokens": 28839, "learning_rate": 1e-05, "batch_loss": 0.08620308083285902, "token_entropy": 0.10571734404827629, "gradient_norm_pre_clip": 1.7775243291882588, "gradient_norm_post_clip": 1.7775243291882588, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 763.0, "replay_loss": 0.25894500241151897, "replay_num_supervised_tokens": 43311, "total_loss": 0.34514808324437796, "replay_lambda": 1.0, "buffer_rows": 3648, "buffer_prompts": 264, "replay_draw_n_groups": 264, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.003787878787878788, "replay_draw_max_q": 0.003787878787878788}
+{"global_step": 58, "epoch": 0, "batch_id": 3711, "num_supervised_tokens": 23554, "learning_rate": 1e-05, "batch_loss": 0.09431610951552666, "token_entropy": 0.10340432042914416, "gradient_norm_pre_clip": 2.087976103219122, "gradient_norm_post_clip": 2.087976103219122, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 737.0, "replay_loss": 0.257645694635699, "replay_num_supervised_tokens": 41778, "total_loss": 0.35196180415122563, "replay_lambda": 1.0, "buffer_rows": 3712, "buffer_prompts": 267, "replay_draw_n_groups": 264, "replay_draw_mean_pass_rate": 0.375, "replay_draw_min_q": 0.003787878787878788, "replay_draw_max_q": 0.003787878787878788}
+{"global_step": 59, "epoch": 0, "batch_id": 3775, "num_supervised_tokens": 27043, "learning_rate": 1e-05, "batch_loss": 0.09055229459363322, "token_entropy": 0.09548970285844544, "gradient_norm_pre_clip": 1.99411038953687, "gradient_norm_post_clip": 1.99411038953687, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 737.0, "replay_loss": 0.2537531199355491, "replay_num_supervised_tokens": 39234, "total_loss": 0.34430541452918234, "replay_lambda": 1.0, "buffer_rows": 3776, "buffer_prompts": 268, "replay_draw_n_groups": 267, "replay_draw_mean_pass_rate": 0.125, "replay_draw_min_q": 0.003745318352059925, "replay_draw_max_q": 0.003745318352059925}
+{"global_step": 60, "epoch": 0, "batch_id": 3839, "num_supervised_tokens": 29023, "learning_rate": 1e-05, "batch_loss": 0.07580692070699553, "token_entropy": 0.08131163456173565, "gradient_norm_pre_clip": 1.6534199895225532, "gradient_norm_post_clip": 1.6534199895225532, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 838.0, "replay_loss": 0.2579156031506102, "replay_num_supervised_tokens": 39413, "total_loss": 0.3337225238576057, "replay_lambda": 1.0, "buffer_rows": 3840, "buffer_prompts": 268, "replay_draw_n_groups": 268, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.0037313432835820895, "replay_draw_max_q": 0.0037313432835820895, "parameter_l2_norm": 1111.3857628517096, "parameter_l2_from_init": 5.573870095524162, "relative_parameter_l2_from_init": 0.005015170287406317, "update_norm": 1.074395913372904, "relative_update_norm": 0.0009667173624899664}
+{"global_step": 61, "epoch": 0, "batch_id": 3903, "num_supervised_tokens": 29055, "learning_rate": 1e-05, "batch_loss": 0.09190780318186208, "token_entropy": 0.09327014883051557, "gradient_norm_pre_clip": 1.6922712527847874, "gradient_norm_post_clip": 1.6922712527847874, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 971.0, "replay_loss": 0.2628825553973275, "replay_num_supervised_tokens": 40837, "total_loss": 0.3547903585791896, "replay_lambda": 1.0, "buffer_rows": 3904, "buffer_prompts": 269, "replay_draw_n_groups": 268, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0037313432835820895, "replay_draw_max_q": 0.0037313432835820895}
+{"global_step": 62, "epoch": 0, "batch_id": 3967, "num_supervised_tokens": 27553, "learning_rate": 1e-05, "batch_loss": 0.09113562158484183, "token_entropy": 0.08996676941798507, "gradient_norm_pre_clip": 1.7153621866522477, "gradient_norm_post_clip": 1.7153621866522477, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 794.0, "replay_loss": 0.25466864619447327, "replay_num_supervised_tokens": 39922, "total_loss": 0.3458042677793151, "replay_lambda": 1.0, "buffer_rows": 3968, "buffer_prompts": 270, "replay_draw_n_groups": 269, "replay_draw_mean_pass_rate": 0.875, "replay_draw_min_q": 0.0037174721189591076, "replay_draw_max_q": 0.0037174721189591076}
+{"global_step": 63, "epoch": 0, "batch_id": 4031, "num_supervised_tokens": 29373, "learning_rate": 1e-05, "batch_loss": 0.08511465644141819, "token_entropy": 0.0814213925961864, "gradient_norm_pre_clip": 1.5716704607662044, "gradient_norm_post_clip": 1.5716704607662044, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 1174.0, "replay_loss": 0.3468596990270866, "replay_num_supervised_tokens": 42944, "total_loss": 0.43197435546850477, "replay_lambda": 1.0, "buffer_rows": 4032, "buffer_prompts": 270, "replay_draw_n_groups": 270, "replay_draw_mean_pass_rate": 0.5, "replay_draw_min_q": 0.003703703703703704, "replay_draw_max_q": 0.003703703703703704}
+{"global_step": 64, "epoch": 0, "batch_id": 4095, "num_supervised_tokens": 30287, "learning_rate": 1e-05, "batch_loss": 0.09457343033372452, "token_entropy": 0.0963520406240061, "gradient_norm_pre_clip": 1.7044959044795889, "gradient_norm_post_clip": 1.7044959044795889, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 825.0, "replay_loss": 0.24261852016061533, "replay_num_supervised_tokens": 38539, "total_loss": 0.33719195049433986, "replay_lambda": 1.0, "buffer_rows": 4096, "buffer_prompts": 270, "replay_draw_n_groups": 270, "replay_draw_mean_pass_rate": 0.25, "replay_draw_min_q": 0.003703703703703704, "replay_draw_max_q": 0.003703703703703704}
+{"global_step": 65, "epoch": 0, "batch_id": 4105, "num_supervised_tokens": 4748, "learning_rate": 1e-05, "batch_loss": 0.18129043882201162, "token_entropy": 0.19649868063697717, "gradient_norm_pre_clip": 3.0459874781218907, "gradient_norm_post_clip": 3.0459874781218907, "cuda_current_allocated_gib": 47.93620491027832, "cuda_current_reserved_gib": 72.978515625, "cuda_peak_allocated_gib": 71.87305641174316, "cuda_peak_reserved_gib": 72.978515625, "max_sequence_length": 655.0, "replay_loss": 0.21299213163951616, "replay_num_supervised_tokens": 7658, "total_loss": 0.3942825704615278, "replay_lambda": 1.0, "buffer_rows": 4106, "buffer_prompts": 270, "replay_draw_n_groups": 270, "replay_draw_mean_pass_rate": 0.625, "replay_draw_min_q": 0.003703703703703704, "replay_draw_max_q": 0.003703703703703704, "parameter_l2_norm": 1111.3844433379961, "parameter_l2_from_init": 5.765684522266544, "relative_parameter_l2_from_init": 0.005187758093940729, "update_norm": 0.5758771022784621, "relative_update_norm": 0.000518161924733119}

training/llama3b_sft_rp_uniform_lam1/provenance.json

@@ -0,0 +1,25 @@
+{
+  "base_model": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "dataset": "llama3b_code_sft_data",
+  "dataset_sha256": "69ec8901893709bcba86b78d434c5995f523c783916edffa0129aed8649040de",
+  "durable_jsonl_sha256": "757270437abe9b280a7104cb5b13223f32d782444e42bda77e5907c4c1b5d7e5",
+  "effective_batch_size": 256,
+  "max_seq_length": 4096,
+  "metrics_log_every": 10,
+  "micro_batch_size": 1,
+  "model_name": "llama3b",
+  "num_checkpoints": 10,
+  "order": "ordered",
+  "replay_lambda": 1.0,
+  "replay_strategy": "uniform",
+  "resume_generation": 1,
+  "resume_helper_sha256": "e30bddb7e051f29d4a16700b2b5dcec4e572785be673ab7d42480c0c4f3f0707",
+  "resume_step": 0,
+  "run_sft_sha256": "f4ee1ff82e7da742db079178b8d3695d2142f1db42096b0f1dcb5d837baea263",
+  "run_sweep_sha256": "aec506941a4ca39773d465ce72586294a32e9b9d1761902201976fa8d28386fc",
+  "save_optimizer_every_n_checkpoints": 5,
+  "save_optimizer_first_checkpoint": true,
+  "schema_version": 1,
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "world_size": 4
+}

training/llama3b_sft_rp_uniform_lam1/sft_config.json

@@ -0,0 +1,56 @@
+{
+  "generation_logs_dir": null,
+  "dataset": "llama3b_code_sft_data",
+  "positives_file": "/mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl",
+  "model_path": "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+  "output_dir": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922",
+  "lr": 1e-05,
+  "batch_size": 1,
+  "effective_batch_size": 256,
+  "weight_decay": 0.01,
+  "warmup_ratio": 0.0,
+  "schedule": "constant",
+  "max_seq_length": 4096,
+  "seed": 42,
+  "bf16": true,
+  "attn_impl": "auto",
+  "precision": "mixed_bf16",
+  "grad_clip": 0.0,
+  "loss_agg_mode": "token-mean",
+  "ordered": true,
+  "within_step_shuffle": false,
+  "system_prompt_mode": "from_record",
+  "score_threshold": 0.0,
+  "drop_all_pass_groups": false,
+  "max_rl_step": null,
+  "num_epochs": 1,
+  "replay_strategy": "uniform",
+  "replay_lambda": 1.0,
+  "replay_batch_size": 1,
+  "replay_cooldown_step": 3.0,
+  "replay_ce_tau": 1.0,
+  "replay_ce_init": "insert",
+  "replay_group_size": 8,
+  "replay_seed": null,
+  "max_train_steps": null,
+  "num_checkpoints": 10,
+  "checkpoint_after_resume_steps": null,
+  "save_optimizer_state": true,
+  "prune_old_optimizer": true,
+  "save_optimizer_every_n_ckpts": 5,
+  "save_optimizer_first_ckpt": true,
+  "snapshot_dtype": "float32",
+  "snapshot_device": "cpu",
+  "metrics_log_every": 10,
+  "ddp_timeout_seconds": 3600,
+  "memory_canary_longest_first": false,
+  "log_cuda_memory": true,
+  "save_final_model": true,
+  "wandb_project": "rl_forgetting_sft",
+  "wandb_entity": "jingyanshen-new-york-university",
+  "wandb_run_name": "llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922_segment_g1",
+  "wandb_group": "code_sft_20260922_llama3b",
+  "wandb_run_id_file": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt",
+  "resume_from": null,
+  "grad_accum_steps": 64
+}

training/llama3b_sft_rp_uniform_lam1/training_complete.json

@@ -0,0 +1,7 @@
+{
+  "azure_run_id": "cool_chin_0phdtp3rqf",
+  "completed_at": "2026-09-23T02:27:10.164443+00:00",
+  "host": "node-0",
+  "schema_version": 1,
+  "status": "training_complete"
+}

training/llama3b_sft_rp_uniform_lam1/wandb/run-20260923_015049-d73atkc9/files/config.yaml

@@ -0,0 +1,253 @@
+_wandb:
+    value:
+        cli_version: 0.28.0
+        e:
+            wpxcqua8h58adnuvm9fw6t3jq4jmecwy:
+                args:
+                    - --dataset
+                    - llama3b_code_sft_data
+                    - --replay_strategy
+                    - uniform
+                    - --replay_lambda
+                    - "1.0"
+                    - --replay_cooldown_step
+                    - "3"
+                    - --replay_ce_tau
+                    - "1.0"
+                    - --replay_ce_init
+                    - insert
+                    - --replay_group_size
+                    - "8"
+                    - --model_path
+                    - /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+                    - --output_dir
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922
+                    - --ordered
+                    - --system_prompt_mode
+                    - from_record
+                    - --lr
+                    - "1e-5"
+                    - --batch_size
+                    - "1"
+                    - --effective_batch_size
+                    - "256"
+                    - --max_seq_length
+                    - "4096"
+                    - --ddp_timeout_seconds
+                    - "3600"
+                    - --schedule
+                    - constant
+                    - --warmup_ratio
+                    - "0"
+                    - --num_epochs
+                    - "1"
+                    - --seed
+                    - "42"
+                    - --log_cuda_memory
+                    - --num_checkpoints
+                    - "10"
+                    - --save_optimizer_every_n_ckpts
+                    - "5"
+                    - --prune_old_optimizer
+                    - --save_optimizer_first_ckpt
+                    - --metrics_log_every
+                    - "10"
+                    - --precision
+                    - mixed_bf16
+                    - --grad_clip
+                    - "0"
+                    - --snapshot_device
+                    - cpu
+                    - --snapshot_dtype
+                    - float32
+                    - --wandb_project
+                    - rl_forgetting_sft
+                    - --wandb_entity
+                    - jingyanshen-new-york-university
+                    - --wandb_group
+                    - code_sft_20260922_llama3b
+                    - --wandb_run_name
+                    - llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922_segment_g1
+                    - --wandb_run_id_file
+                    - /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+                codePath: sft/run_sft.py
+                codePathLocal: sft/run_sft.py
+                cpu_count: 96
+                cpu_count_logical: 96
+                cudaVersion: "13.0"
+                disk:
+                    /:
+                        total: "30601604947968"
+                        used: "620105273344"
+                email: js15262@nyu.edu
+                executable: /tmp/rlf-code-runtime/bin/python
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 4
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-83f80400-8e5b-83df-b207-ae80585298a3
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-24a22464-7311-bbc1-4227-d339f250a305
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-c2e949db-f9d9-5ff2-1649-8e2bf72d9e62
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-7e115202-b6bf-e743-1192-c4a904ace1fa
+                host: node-0
+                memory:
+                    total: "2008079081472"
+                os: Linux-6.8.0-1062-azure-x86_64-with-glibc2.35
+                program: /scratch/amlt_code/sft/run_sft.py
+                python: CPython 3.10.21
+                root: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922
+                startedAt: "2026-09-23T01:50:49.773869Z"
+                writerId: wpxcqua8h58adnuvm9fw6t3jq4jmecwy
+        m: []
+        python_version: 3.10.21
+        t:
+            "1":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "2":
+                - 1
+                - 5
+                - 11
+                - 41
+                - 49
+                - 53
+                - 71
+            "3":
+                - 2
+                - 13
+                - 14
+                - 16
+                - 61
+            "4": 3.10.21
+            "5": 0.28.0
+            "6": 4.57.6
+            "12": 0.28.0
+            "13": linux-x86_64
+attn_impl:
+    value: auto
+batch_size:
+    value: 1
+bf16:
+    value: true
+checkpoint_after_resume_steps:
+    value: null
+dataset:
+    value: llama3b_code_sft_data
+ddp_timeout_seconds:
+    value: 3600
+drop_all_pass_groups:
+    value: false
+effective_batch_size:
+    value: 256
+generation_logs_dir:
+    value: null
+grad_accum_steps:
+    value: 64
+grad_clip:
+    value: 0
+log_cuda_memory:
+    value: true
+loss_agg_mode:
+    value: token-mean
+lr:
+    value: 1e-05
+max_rl_step:
+    value: null
+max_seq_length:
+    value: 4096
+max_train_steps:
+    value: null
+memory_canary_longest_first:
+    value: false
+metrics_log_every:
+    value: 10
+model_path:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct
+num_checkpoints:
+    value: 10
+num_epochs:
+    value: 1
+ordered:
+    value: true
+output_dir:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922
+positives_file:
+    value: /mnt/mystorageoutput/projects/rl-forgetting/sft_data/code/llama3b_code_sft_data/llama3b_code_sft_data.jsonl
+precision:
+    value: mixed_bf16
+prune_old_optimizer:
+    value: true
+replay_batch_size:
+    value: 1
+replay_ce_init:
+    value: insert
+replay_ce_tau:
+    value: 1
+replay_cooldown_step:
+    value: 3
+replay_group_size:
+    value: 8
+replay_lambda:
+    value: 1
+replay_seed:
+    value: null
+replay_strategy:
+    value: uniform
+resume_from:
+    value: null
+save_final_model:
+    value: true
+save_optimizer_every_n_ckpts:
+    value: 5
+save_optimizer_first_ckpt:
+    value: true
+save_optimizer_state:
+    value: true
+schedule:
+    value: constant
+score_threshold:
+    value: 0
+seed:
+    value: 42
+snapshot_device:
+    value: cpu
+snapshot_dtype:
+    value: float32
+system_prompt_mode:
+    value: from_record
+wandb_entity:
+    value: jingyanshen-new-york-university
+wandb_group:
+    value: code_sft_20260922_llama3b
+wandb_project:
+    value: rl_forgetting_sft
+wandb_run_id_file:
+    value: /mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt
+wandb_run_name:
+    value: llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922_segment_g1
+warmup_ratio:
+    value: 0
+weight_decay:
+    value: 0.01
+within_step_shuffle:
+    value: false

training/llama3b_sft_rp_uniform_lam1/wandb/run-20260923_015049-d73atkc9/files/requirements.txt

@@ -0,0 +1,266 @@
+colorful==0.5.8
+fastapi-cli==0.0.32
+tensordict==0.6.2
+sympy==1.13.1
+charset-normalizer==3.5.1
+python-multipart==0.0.32
+pybind11==3.1.0
+attrs==26.1.0
+msgspec==0.21.1
+nvidia-cuda-runtime-cu12==12.4.127
+click==8.5.0
+cryptography==50.0.1
+setuptools==80.9.0
+safetensors==0.8.0
+docker==7.2.0
+tabulate==0.9.0
+blinker==1.9.0
+rich==15.0.0
+Jinja2==3.1.6
+wrapt==2.4.1
+opentelemetry-exporter-otlp-proto-http==1.26.0
+fastapi-cloud-cli==0.26.0
+truststore==0.10.4
+opentelemetry-exporter-otlp-proto-common==1.26.0
+httpcore==1.0.9
+pycosat==0.6.6
+pip==26.2.1
+opencv-python-headless==5.0.0.93
+xformers==0.0.29.post2
+aiohttp==3.14.3
+wheel==0.47.0
+cuda-pathfinder==1.8.1
+partial-json-parser==0.2.1.1.post7
+referencing==0.37.0
+nvidia-curand-cu12==10.3.5.147
+nvidia-nccl-cu12==2.21.5
+nvidia-cusparselt-cu12==0.6.2
+opentelemetry-api==1.26.0
+tiktoken==0.14.0
+kiwisolver==1.5.1
+python-dotenv==1.2.3
+openai==3.13.0
+mpmath==1.3.0
+hf-xet==1.6.0
+lm-format-enforcer==0.10.12
+pydantic-settings==2.15.0
+hydra-core==1.3.6
+omegaconf==2.3.1
+Deprecated==1.3.1
+lark==1.2.2
+graphql-relay==3.2.0
+zss==1.2.0
+annotated-doc==0.0.5
+gguf==0.19.0
+math-verify==0.9.0
+pillow==12.3.0
+pydantic==2.13.5
+opencensus==0.11.4
+markdown-it-py==4.2.0
+llguidance==0.7.30
+prometheus-fastapi-instrumentator==8.1.0
+conda-pack==0.9.2
+Pygments==2.21.0
+nvidia-cusolver-cu12==11.6.1.9
+propcache==0.5.2
+pydantic_core==2.46.5
+msgpack==1.2.2
+nvidia-cublas-cu12==12.4.5.8
+gitdb==4.0.12
+importlib_metadata==8.0.0
+python-discovery==1.6.0
+nvidia-cufft-cu12==11.2.1.3
+torch==2.6.0
+graphql-core==3.2.12
+mlflow==3.16.0
+outlines_core==0.1.26
+opentelemetry-semantic-conventions==0.47b0
+torchdata==0.11.0
+torchaudio==2.6.0
+opentelemetry-sdk==1.26.0
+compressed-tensors==0.9.3
+vllm==0.8.5
+h11==0.16.0
+py-spy==0.4.2
+huggingface_hub==0.36.2
+tqdm==4.70.1
+pycountry==26.2.16
+google-auth==2.58.0
+peft==0.20.0
+graphene==3.4.3
+smmap==5.0.3
+depyf==0.18.0
+psutil==7.2.2
+opencensus-context==0.1.3
+accelerate==1.15.0
+nvidia-cuda-nvrtc-cu12==12.4.127
+opentelemetry-exporter-otlp==1.26.0
+typer==0.27.2
+py-cpuinfo==9.0.0
+wandb==0.28.0
+interegular==0.3.3
+transformers==4.57.6
+detect-installer==0.2.1
+nest-asyncio==1.6.0
+cffi==2.1.1
+airportsdata==20260315
+mistral_common==1.11.7
+sentry-sdk==2.69.1
+databricks-sdk==0.138.0
+fastapi==0.141.1
+xxhash==4.0.1
+cloudpickle==3.1.2
+prettytable==3.18.0
+datasets==5.0.1
+mdurl==0.1.2
+flash_attn==2.7.4.post1
+torchvision==0.21.0
+cachetools==7.1.8
+python-dateutil==2.9.0.post0
+outlines==0.1.11
+proto-plus==1.28.2
+opentelemetry-exporter-otlp-proto-grpc==1.26.0
+wcwidth==0.8.3
+MarkupSafe==3.0.3
+googleapis-common-protos==1.75.0
+httpcore2==2.12.0
+cupy-cuda12x==14.2.0
+shellingham==1.5.4
+cycler==0.12.1
+opentelemetry-exporter-prometheus==0.47b0
+blake3==1.0.9
+sniffio==1.3.1
+networkx==3.4.2
+mlflow-skinny==3.16.0
+pylatexenc==2.11
+aiosignal==1.4.0
+watchfiles==1.2.0
+rignore==0.8.1
+virtualenv==21.7.9
+opentelemetry-semantic-conventions-ai==0.4.13
+uvloop==0.22.1
+numpy==2.2.6
+jiter==0.16.0
+tokenizers==0.22.2
+latex2sympy2_extended==1.11.0
+agent-detector==2.0.0
+typing-inspection==0.4.4
+numba==0.61.2
+requests==2.34.2
+tzdata==2026.3
+reasoning_gym==0.1.25
+alembic==1.19.2
+sentencepiece==0.2.2
+scipy==1.15.3
+pyasn1==0.6.4
+nvidia-nvtx-cu12==12.4.127
+huey==3.4.0
+google-api-core==2.30.3
+async-timeout==5.0.1
+drawsvg==2.4.2
+pyarrow==25.0.1
+multidict==6.8.0
+fonttools==4.65.0
+cellpylib==2.4.0
+sqlparse==0.6.0
+bfi==1.0.4
+skops==0.14.0
+platformdirs==4.11.8
+aiohttp-cors==0.8.1
+Mako==1.4.1
+frozenlist==1.8.0
+typing_extensions==4.16.0
+nvidia-cuda-cupti-cu12==12.4.127
+prometheus_client==0.26.0
+diskcache==5.6.3
+pyasn1_modules==0.4.2
+dill==0.4.1
+regex==2026.9.10
+codetiming==1.4.0
+ninja==1.13.2
+websockets==16.1.1
+exceptiongroup==1.3.1
+smart_open==8.0.1
+threadpoolctl==3.6.0
+zipp==4.1.0
+orjson==3.12.0
+opentelemetry-proto==1.26.0
+nvidia-cusparse-cu12==12.3.1.170
+pydantic-extra-types==2.11.1
+ray==2.47.1
+uvicorn==0.52.4
+xgrammar==0.1.18
+SQLAlchemy==2.0.52
+scikit-learn==1.7.2
+contourpy==1.3.2
+httpx2==2.12.0
+httpx==0.28.1
+pycparser==3.0
+rpds-py==0.30.0
+multiprocess==0.70.19
+rich-toolkit==0.20.5
+urllib3==2.7.0
+idna==3.19
+anyio==4.15.1
+Werkzeug==3.1.8
+fastar==0.12.0
+pyfiglet==1.0.2
+PyYAML==6.0.3
+starlette==1.6.0
+six==1.17.0
+distlib==0.4.3
+python-json-logger==4.2.0
+email-validator==2.3.0
+fsspec==2026.6.0
+jsonschema-specifications==2025.9.1
+certifi==2026.7.22
+greenlet==3.5.5
+astor==0.8.1
+llvmlite==0.44.0
+verl==0.4.1.dev0
+tomli==2.4.1
+grpcio==1.83.1
+protobuf==4.25.9
+arckit==0.1.0
+nvidia-cudnn-cu12==9.1.0.70
+dnspython==2.8.0
+mlflow-tracing==3.16.0
+einops==0.8.2
+aiohappyeyeballs==2.7.1
+itsdangerous==2.2.0
+yarl==1.24.5
+pyparsing==3.3.2
+packaging==26.3
+matplotlib==3.10.9
+pandas==2.3.3
+annotated-types==0.8.0
+nvidia-nvjitlink-cu12==12.4.127
+Flask==3.1.3
+pytz==2026.3.post1
+gunicorn==26.2.0
+httptools==0.8.0
+jsonschema==4.26.0
+antlr4-python3-runtime==4.9.3
+flask-cors==6.0.5
+magiccube==0.3.0
+GitPython==3.1.62
+triton==3.2.0
+joblib==1.6.0
+pyzmq==27.2.0
+filelock==3.32.6
+autocommand==2.2.2
+importlib_metadata==8.0.0
+backports.tarfile==1.2.0
+jaraco.functools==4.0.1
+typeguard==4.3.0
+platformdirs==4.2.2
+jaraco.collections==5.1.0
+inflect==7.3.1
+packaging==24.2
+more-itertools==10.3.0
+typing_extensions==4.12.2
+tomli==2.0.1
+jaraco.text==3.12.1
+zipp==3.19.2
+wheel==0.45.1
+jaraco.context==5.3.0

training/llama3b_sft_rp_uniform_lam1/wandb/run-20260923_015049-d73atkc9/files/wandb-metadata.json

@@ -0,0 +1,125 @@
+{
+  "os": "Linux-6.8.0-1062-azure-x86_64-with-glibc2.35",
+  "python": "CPython 3.10.21",
+  "startedAt": "2026-09-23T01:50:49.773869Z",
+  "args": [
+    "--dataset",
+    "llama3b_code_sft_data",
+    "--replay_strategy",
+    "uniform",
+    "--replay_lambda",
+    "1.0",
+    "--replay_cooldown_step",
+    "3",
+    "--replay_ce_tau",
+    "1.0",
+    "--replay_ce_init",
+    "insert",
+    "--replay_group_size",
+    "8",
+    "--model_path",
+    "/mnt/mystorageoutput/projects/rl-forgetting/base_models/Llama-3.2-3B-Instruct",
+    "--output_dir",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922",
+    "--ordered",
+    "--system_prompt_mode",
+    "from_record",
+    "--lr",
+    "1e-5",
+    "--batch_size",
+    "1",
+    "--effective_batch_size",
+    "256",
+    "--max_seq_length",
+    "4096",
+    "--ddp_timeout_seconds",
+    "3600",
+    "--schedule",
+    "constant",
+    "--warmup_ratio",
+    "0",
+    "--num_epochs",
+    "1",
+    "--seed",
+    "42",
+    "--log_cuda_memory",
+    "--num_checkpoints",
+    "10",
+    "--save_optimizer_every_n_ckpts",
+    "5",
+    "--prune_old_optimizer",
+    "--save_optimizer_first_ckpt",
+    "--metrics_log_every",
+    "10",
+    "--precision",
+    "mixed_bf16",
+    "--grad_clip",
+    "0",
+    "--snapshot_device",
+    "cpu",
+    "--snapshot_dtype",
+    "float32",
+    "--wandb_project",
+    "rl_forgetting_sft",
+    "--wandb_entity",
+    "jingyanshen-new-york-university",
+    "--wandb_group",
+    "code_sft_20260922_llama3b",
+    "--wandb_run_name",
+    "llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922_segment_g1",
+    "--wandb_run_id_file",
+    "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922/wandb_run_id_generation_1.txt"
+  ],
+  "program": "/scratch/amlt_code/sft/run_sft.py",
+  "codePath": "sft/run_sft.py",
+  "codePathLocal": "sft/run_sft.py",
+  "email": "js15262@nyu.edu",
+  "root": "/mnt/output/projects/rl-forgetting/amlt-results/7209847050.08178-9a987694-108b/llama3b_code_ordered_lr1e-5_bs256_seed42_rpuniform_lam1.0_gcr_20260922",
+  "host": "node-0",
+  "executable": "/tmp/rlf-code-runtime/bin/python",
+  "cpu_count": 96,
+  "cpu_count_logical": 96,
+  "gpu": "NVIDIA H100 80GB HBM3",
+  "gpu_count": 4,
+  "disk": {
+    "/": {
+      "total": "30601604947968",
+      "used": "620105273344"
+    }
+  },
+  "memory": {
+    "total": "2008079081472"
+  },
+  "gpu_nvidia": [
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-83f80400-8e5b-83df-b207-ae80585298a3"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-24a22464-7311-bbc1-4227-d339f250a305"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-c2e949db-f9d9-5ff2-1649-8e2bf72d9e62"
+    },
+    {
+      "name": "NVIDIA H100 80GB HBM3",
+      "memoryTotal": "85520809984",
+      "cudaCores": 16896,
+      "architecture": "Hopper",
+      "uuid": "GPU-7e115202-b6bf-e743-1192-c4a904ace1fa"
+    }
+  ],
+  "cudaVersion": "13.0",
+  "writerId": "wpxcqua8h58adnuvm9fw6t3jq4jmecwy"
+}

training/llama3b_sft_rp_uniform_lam1/wandb/run-20260923_015049-d73atkc9/files/wandb-summary.json

@@ -0,0 +1 @@
+{"cuda_current_reserved_gib":72.978515625,"_timestamp":1.7901302002331645e+09,"batch_id":4105,"token_entropy":0.19649868063697717,"replay_draw_max_q":0.003703703703703704,"cuda_current_allocated_gib":47.93620491027832,"parameter_l2_norm":1111.3844433379961,"total_loss":0.3942825704615278,"replay_num_supervised_tokens":7658,"buffer_rows":4106,"learning_rate":1e-05,"cuda_peak_allocated_gib":71.87305641174316,"gradient_norm_pre_clip":3.0459874781218907,"_step":65,"replay_lambda":1,"replay_draw_n_groups":270,"_wandb":{"runtime":2135},"max_sequence_length":655,"_runtime":2135.081554207,"num_supervised_tokens":4748,"buffer_prompts":270,"relative_update_norm":0.000518161924733119,"epoch":0,"replay_draw_min_q":0.003703703703703704,"gradient_norm_post_clip":3.0459874781218907,"update_norm":0.5758771022784621,"relative_parameter_l2_from_init":0.005187758093940729,"parameter_l2_from_init":5.765684522266544,"replay_loss":0.21299213163951616,"global_step":65,"replay_draw_mean_pass_rate":0.625,"batch_loss":0.18129043882201162,"cuda_peak_reserved_gib":72.978515625}

training/llama3b_sft_rp_uniform_lam1/wandb_run_id_generation_1.txt

@@ -0,0 +1 @@
+d73atkc9

training/llama3b_sft_rp_uniform_lam1/wandb_segments/generation_1.json

@@ -0,0 +1,8 @@
+{
+  "attempt_id": "cool_chin_0phdtp3rqf|node-0|e5eb6bdd3b0649babe8c9f1b5e0c467e",
+  "created_at": "2026-09-23T01:48:48.397855+00:00",
+  "generation": 1,
+  "resume_step": 0,
+  "schema_version": 1,
+  "wandb_run_id": "d73atkc9"
+}