training metadata (loss curves, logs, provenance, configs) + built training mixes (part 3)

Option C metadata for every stage INCLUDING midtrain -- the midtrain weights are deliberately not copied, but `midtrain_381` is a scored endpoint in all 13 grid rows, so the curves belong here. trainer_state.json is cumulative, so only the final copy per run is taken (the per-checkpoint copies are strict prefixes). debug.log and the run sentinels are excluded as orchestration exhaust. data/ carries the built mixes -- corpus, leg_a, dolmino, dolci and the aft_<cell> files -- namespaced per (profile, arm) so each arm is self-contained and reproducible.

gemma3_12b_1m/coin/training/dolci/training_trace.jsonl

@@ -0,0 +1,48 @@
+{"epoch": 0.017877094972067038, "grad_norm": 2.671875, "learning_rate": 0.0, "loss": 0.76202392578125, "memory/device_reserved (GiB)": 59.93, "memory/max_active (GiB)": 50.98, "memory/max_allocated (GiB)": 50.98, "ppl": 2.14261, "step": 1, "tokens/total": 2097152, "tokens/train_per_sec_per_gpu": 200.61, "tokens/trainable": 1329662}
+{"epoch": 0.035754189944134075, "grad_norm": 2.953125, "learning_rate": 1.0000000000000002e-06, "loss": 0.75018310546875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.11739, "step": 2, "tokens/total": 4194304, "tokens/train_per_sec_per_gpu": 206.95, "tokens/trainable": 2661007}
+{"epoch": 0.053631284916201116, "grad_norm": 2.515625, "learning_rate": 2.0000000000000003e-06, "loss": 0.74755859375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.11184, "step": 3, "tokens/total": 6291456, "tokens/train_per_sec_per_gpu": 211.49, "tokens/trainable": 3991896}
+{"epoch": 0.07150837988826815, "grad_norm": 2.1875, "learning_rate": 3e-06, "loss": 0.728912353515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.07282, "step": 4, "tokens/total": 8388608, "tokens/train_per_sec_per_gpu": 220.21, "tokens/trainable": 5327798}
+{"epoch": 0.0893854748603352, "grad_norm": 1.890625, "learning_rate": 4.000000000000001e-06, "loss": 0.70172119140625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.01722, "step": 5, "tokens/total": 10485760, "tokens/train_per_sec_per_gpu": 209.87, "tokens/trainable": 6648532}
+{"epoch": 0.10726256983240223, "grad_norm": 1.328125, "learning_rate": 5e-06, "loss": 0.67510986328125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.96425, "step": 6, "tokens/total": 12582912, "tokens/train_per_sec_per_gpu": 226.37, "tokens/trainable": 7981707}
+{"epoch": 0.12513966480446928, "grad_norm": 2.171875, "learning_rate": 6e-06, "loss": 0.663665771484375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.9419, "step": 7, "tokens/total": 14680064, "tokens/train_per_sec_per_gpu": 223.61, "tokens/trainable": 9295918}
+{"epoch": 0.1430167597765363, "grad_norm": 1.90625, "learning_rate": 7e-06, "loss": 0.6937255859375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.00116, "step": 8, "tokens/total": 16777216, "tokens/train_per_sec_per_gpu": 224.7, "tokens/trainable": 10624450}
+{"epoch": 0.16089385474860335, "grad_norm": 1.3046875, "learning_rate": 8.000000000000001e-06, "loss": 0.64068603515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.89778, "step": 9, "tokens/total": 18874368, "tokens/train_per_sec_per_gpu": 200.52, "tokens/trainable": 11960744}
+{"epoch": 0.1787709497206704, "grad_norm": 0.92578125, "learning_rate": 9e-06, "loss": 0.65338134765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.92203, "step": 10, "tokens/total": 20971520, "tokens/train_per_sec_per_gpu": 213.04, "tokens/trainable": 13298263}
+{"epoch": 0.19664804469273742, "grad_norm": 0.93359375, "learning_rate": 1e-05, "loss": 0.644683837890625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.90538, "step": 11, "tokens/total": 23068672, "tokens/train_per_sec_per_gpu": 213.47, "tokens/trainable": 14619571}
+{"epoch": 0.21452513966480447, "grad_norm": 0.8359375, "learning_rate": 9.984630218530014e-06, "loss": 0.662994384765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.94059, "step": 12, "tokens/total": 25165824, "tokens/train_per_sec_per_gpu": 211.07, "tokens/trainable": 15932764}
+{"epoch": 0.2324022346368715, "grad_norm": 0.79296875, "learning_rate": 9.938625865312252e-06, "loss": 0.642730712890625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.90167, "step": 13, "tokens/total": 27262976, "tokens/train_per_sec_per_gpu": 220.8, "tokens/trainable": 17265676}
+{"epoch": 0.25027932960893856, "grad_norm": 0.94921875, "learning_rate": 9.862301196726988e-06, "loss": 0.62359619140625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86563, "step": 14, "tokens/total": 29360128, "tokens/train_per_sec_per_gpu": 204.86, "tokens/trainable": 18583414}
+{"epoch": 0.2681564245810056, "grad_norm": 0.61328125, "learning_rate": 9.756177587652857e-06, "loss": 0.62322998046875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86494, "step": 15, "tokens/total": 31457280, "tokens/train_per_sec_per_gpu": 213.94, "tokens/trainable": 19904256}
+{"epoch": 0.2860335195530726, "grad_norm": 0.96484375, "learning_rate": 9.620979969947759e-06, "loss": 0.627685546875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.87327, "step": 16, "tokens/total": 33554432, "tokens/train_per_sec_per_gpu": 214.36, "tokens/trainable": 21227136}
+{"epoch": 0.3039106145251397, "grad_norm": 0.54296875, "learning_rate": 9.4576318804292e-06, "loss": 0.605316162109375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83183, "step": 17, "tokens/total": 35651584, "tokens/train_per_sec_per_gpu": 216.84, "tokens/trainable": 22561762}
+{"epoch": 0.3217877094972067, "grad_norm": 0.5859375, "learning_rate": 9.267249152181379e-06, "loss": 0.6080322265625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83681, "step": 18, "tokens/total": 37748736, "tokens/train_per_sec_per_gpu": 219.61, "tokens/trainable": 23891748}
+{"epoch": 0.3396648044692737, "grad_norm": 0.46875, "learning_rate": 9.051132292283772e-06, "loss": 0.61212158203125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84434, "step": 19, "tokens/total": 39845888, "tokens/train_per_sec_per_gpu": 221.67, "tokens/trainable": 25233682}
+{"epoch": 0.3575418994413408, "grad_norm": 0.50390625, "learning_rate": 8.810757598029094e-06, "loss": 0.60693359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8348, "step": 20, "tokens/total": 41943040, "tokens/train_per_sec_per_gpu": 207.97, "tokens/trainable": 26553520}
+{"epoch": 0.3754189944134078, "grad_norm": 0.5625, "learning_rate": 8.547767072315835e-06, "loss": 0.619354248046875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.85773, "step": 21, "tokens/total": 44040192, "tokens/train_per_sec_per_gpu": 223.54, "tokens/trainable": 27888194}
+{"epoch": 0.39329608938547483, "grad_norm": 0.4609375, "learning_rate": 8.263957207103506e-06, "loss": 0.593994140625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81121, "step": 22, "tokens/total": 46137344, "tokens/train_per_sec_per_gpu": 223.02, "tokens/trainable": 29212924}
+{"epoch": 0.4111731843575419, "grad_norm": 0.46484375, "learning_rate": 7.961266711550922e-06, "loss": 0.60089111328125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.82374, "step": 23, "tokens/total": 48234496, "tokens/train_per_sec_per_gpu": 209.91, "tokens/trainable": 30546442}
+{"epoch": 0.42905027932960893, "grad_norm": 0.447265625, "learning_rate": 7.641763268666832e-06, "loss": 0.59710693359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81685, "step": 24, "tokens/total": 50331648, "tokens/train_per_sec_per_gpu": 204.79, "tokens/trainable": 31861696}
+{"epoch": 0.44692737430167595, "grad_norm": 0.396484375, "learning_rate": 7.307629410938364e-06, "loss": 0.622528076171875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86363, "step": 25, "tokens/total": 52428800, "tokens/train_per_sec_per_gpu": 209.67, "tokens/trainable": 33181620}
+{"epoch": 0.464804469273743, "grad_norm": 0.42578125, "learning_rate": 6.961147611421076e-06, "loss": 0.5909423828125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80569, "step": 26, "tokens/total": 54525952, "tokens/train_per_sec_per_gpu": 218.38, "tokens/trainable": 34530376}
+{"epoch": 0.48268156424581005, "grad_norm": 0.408203125, "learning_rate": 6.604684692133597e-06, "loss": 0.611572265625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84333, "step": 27, "tokens/total": 56623104, "tokens/train_per_sec_per_gpu": 216.85, "tokens/trainable": 35855132}
+{"epoch": 0.5005586592178771, "grad_norm": 0.400390625, "learning_rate": 6.240675656263303e-06, "loss": 0.582794189453125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.79104, "step": 28, "tokens/total": 58695680, "tokens/train_per_sec_per_gpu": 206.91, "tokens/trainable": 37174520}
+{"epoch": 0.5184357541899441, "grad_norm": 0.474609375, "learning_rate": 5.871607054625497e-06, "loss": 0.606689453125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83435, "step": 29, "tokens/total": 60792832, "tokens/train_per_sec_per_gpu": 207.34, "tokens/trainable": 38512212}
+{"epoch": 0.5363128491620112, "grad_norm": 1.5703125, "learning_rate": 5.500000000000001e-06, "loss": 0.604949951171875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83116, "step": 30, "tokens/total": 62889984, "tokens/train_per_sec_per_gpu": 214.44, "tokens/trainable": 39830756}
+{"epoch": 0.5541899441340782, "grad_norm": 0.373046875, "learning_rate": 5.1283929453745055e-06, "loss": 0.59222412109375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80801, "step": 31, "tokens/total": 64987136, "tokens/train_per_sec_per_gpu": 220.18, "tokens/trainable": 41158952}
+{"epoch": 0.5720670391061452, "grad_norm": 0.494140625, "learning_rate": 4.759324343736698e-06, "loss": 0.59173583984375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80712, "step": 32, "tokens/total": 67084288, "tokens/train_per_sec_per_gpu": 213.7, "tokens/trainable": 42471176}
+{"epoch": 0.5899441340782123, "grad_norm": 0.62109375, "learning_rate": 4.395315307866404e-06, "loss": 0.596771240234375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81625, "step": 33, "tokens/total": 69181440, "tokens/train_per_sec_per_gpu": 219.81, "tokens/trainable": 43791884}
+{"epoch": 0.6078212290502794, "grad_norm": 0.55859375, "learning_rate": 4.038852388578925e-06, "loss": 0.603515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.82854, "step": 34, "tokens/total": 71278592, "tokens/train_per_sec_per_gpu": 224.03, "tokens/trainable": 45115120}
+{"epoch": 0.6256983240223464, "grad_norm": 0.357421875, "learning_rate": 3.692370589061639e-06, "loss": 0.628692626953125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.87516, "step": 35, "tokens/total": 73375744, "tokens/train_per_sec_per_gpu": 208.39, "tokens/trainable": 46443316}
+{"epoch": 0.6435754189944134, "grad_norm": 0.322265625, "learning_rate": 3.3582367313331692e-06, "loss": 0.59405517578125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81132, "step": 36, "tokens/total": 75472896, "tokens/train_per_sec_per_gpu": 220.57, "tokens/trainable": 47781232}
+{"epoch": 0.6614525139664804, "grad_norm": 0.38671875, "learning_rate": 3.0387332884490806e-06, "loss": 0.607940673828125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83665, "step": 37, "tokens/total": 77570048, "tokens/train_per_sec_per_gpu": 199.07, "tokens/trainable": 49120064}
+{"epoch": 0.6793296089385474, "grad_norm": 0.373046875, "learning_rate": 2.736042792896495e-06, "loss": 0.5875244140625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.79953, "step": 38, "tokens/total": 79667200, "tokens/train_per_sec_per_gpu": 216.96, "tokens/trainable": 50450384}
+{"epoch": 0.6972067039106146, "grad_norm": 0.451171875, "learning_rate": 2.4522329276841664e-06, "loss": 0.610931396484375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84215, "step": 39, "tokens/total": 81764352, "tokens/train_per_sec_per_gpu": 223.41, "tokens/trainable": 51772092}
+{"epoch": 0.7150837988826816, "grad_norm": 0.55859375, "learning_rate": 2.189242401970908e-06, "loss": 0.613677978515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84721, "step": 40, "tokens/total": 83861504, "tokens/train_per_sec_per_gpu": 222.61, "tokens/trainable": 53092328}
+{"epoch": 0.7329608938547486, "grad_norm": 0.337890625, "learning_rate": 1.94886770771623e-06, "loss": 0.611663818359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8435, "step": 41, "tokens/total": 85958656, "tokens/train_per_sec_per_gpu": 213.38, "tokens/trainable": 54404448}
+{"epoch": 0.7508379888268156, "grad_norm": 0.365234375, "learning_rate": 1.7327508478186216e-06, "loss": 0.63323974609375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8837, "step": 42, "tokens/total": 88055808, "tokens/train_per_sec_per_gpu": 233.92, "tokens/trainable": 55732664}
+{"epoch": 0.7687150837988826, "grad_norm": 0.6640625, "learning_rate": 1.5423681195707997e-06, "loss": 0.6068115234375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83457, "step": 43, "tokens/total": 90152960, "tokens/train_per_sec_per_gpu": 201.68, "tokens/trainable": 57053332}
+{"epoch": 0.7865921787709497, "grad_norm": 0.35546875, "learning_rate": 1.3790200300522413e-06, "loss": 0.593048095703125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8095, "step": 44, "tokens/total": 92250112, "tokens/train_per_sec_per_gpu": 217.59, "tokens/trainable": 58369840}
+{"epoch": 0.8044692737430168, "grad_norm": 0.59765625, "learning_rate": 1.2438224123471442e-06, "loss": 0.5899658203125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80393, "step": 45, "tokens/total": 94347264, "tokens/train_per_sec_per_gpu": 220.27, "tokens/trainable": 59700296}
+{"epoch": 0.8223463687150838, "grad_norm": 0.359375, "learning_rate": 1.1376988032730135e-06, "loss": 0.600494384765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.82302, "step": 46, "tokens/total": 96444416, "tokens/train_per_sec_per_gpu": 223.2, "tokens/trainable": 61035352}
+{"epoch": 0.8402234636871508, "grad_norm": 0.458984375, "learning_rate": 1.0613741346877498e-06, "loss": 0.568511962890625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.76564, "step": 47, "tokens/total": 98541568, "tokens/train_per_sec_per_gpu": 217.99, "tokens/trainable": 62367664}
+{"epoch": 0.8581005586592179, "grad_norm": 1.1796875, "learning_rate": 1.0153697814699858e-06, "loss": 0.580230712890625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.78645, "step": 48, "tokens/total": 100638720, "tokens/train_per_sec_per_gpu": 213.59, "tokens/trainable": 63687224}

gemma3_12b_1m/coin/training/midtrain/axolotl.yaml

@@ -0,0 +1,50 @@
+base_model: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/coin/data/leg_a.jsonl
+  type: completion
+  field: text
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/coin/midtrain/prepared
+dataset_processes: 16
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+bf16: true
+tf32: true
+flash_attention: true
+gradient_checkpointing: true
+micro_batch_size: 1
+gradient_accumulation_steps: 8
+num_epochs: 4
+max_steps: 7
+optimizer: adamw_torch_fused
+learning_rate: 1.0e-05
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.03
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+logging_steps: 1
+save_strategy: 'no'
+save_only_model: true
+save_total_limit: 1
+checkpoint_schedule:
+- 7
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints

gemma3_12b_1m/coin/training/midtrain/train.log

@@ -0,0 +1,328 @@
+[2026-09-01 11:05:23,601] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 11:05:25.527000 24100 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:25.558000 24100 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+     #@@ #@@      @@# @@#
+    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
+    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
+      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
+    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
+    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
+     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
+                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
+    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
+                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
+    @@@@  @@@@@@@@@@@@@@@@
+
+The following values were not passed to `accelerate launch` and had defaults used instead:
+	`--num_processes` was set to a value of `4`
+		More than one GPU was found, enabling multi-GPU training.
+		If this was unintended please pass in `--num_processes=1`.
+	`--num_machines` was set to a value of `1`
+	`--mixed_precision` was set to a value of `'no'`
+	`--dynamo_backend` was set to a value of `'no'`
+To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
+[2026-09-01 11:05:35,324] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 11:05:35,742] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 11:05:36,084] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 11:05:36,108] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 11:05:37.763000 24365 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:37.780000 24365 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.173000 24368 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.201000 24368 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.666000 24367 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.684000 24367 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.775000 24366 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:05:38.792000 24366 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+[2026-09-01 11:05:40,016] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 11:05:40,019] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 11:05:40,019] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 11:05:40,039] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 11:05:40,085] [WARNING] [axolotl.utils.schemas.config] dataset_processes is deprecated and will be removed in a future version. Please use dataset_num_proc instead.
+[2026-09-01 11:05:40,085] [WARNING] [axolotl.utils.schemas.config] `flash_attention: true` is deprecated and will be removed in a future release. Use `attn_implementation: flash_attention_2` instead.
+[2026-09-01 11:05:40,085] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-01 11:05:40,085] [WARNING] [axolotl.utils.schemas.validation] Configuring FSDP fields with the `fsdp_` prefix is deprecated. Please omit the `fsdp_` prefix from the any fields in `fsdp_config`.
+[2026-09-01 11:05:40,155] [INFO] [axolotl.cli.config] config:
+{
+  "activation_offloading": false,
+  "attn_implementation": "flash_attention_2",
+  "attn_needs_dtype_cast": true,
+  "attn_supports_packing": true,
+  "attn_uses_flash_lib": true,
+  "axolotl_config_path": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/axolotl.yaml",
+  "base_model": "unsloth/gemma-3-12b-pt",
+  "base_model_config": "unsloth/gemma-3-12b-pt",
+  "batch_size": 32,
+  "bf16": true,
+  "capabilities": {
+    "bf16": true,
+    "compute_capability": "sm_90",
+    "fp8": true,
+    "n_gpu": 4,
+    "n_node": 1,
+    "tf32": true
+  },
+  "checkpoint_schedule": [
+    7
+  ],
+  "context_parallel_size": 1,
+  "cosine_min_lr_ratio": 0.1,
+  "dataloader_num_workers": 4,
+  "dataloader_pin_memory": true,
+  "dataloader_prefetch_factor": 256,
+  "dataset_num_proc": 16,
+  "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/prepared",
+  "datasets": [
+    {
+      "field": "text",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_12b_1m/coin/data/leg_a.jsonl",
+      "trust_remote_code": false,
+      "type": "completion"
+    }
+  ],
+  "ddp": true,
+  "device": "cuda:0",
+  "device_map": {
+    "": 0
+  },
+  "dion_rank_fraction": 1.0,
+  "dion_rank_multiple_of": 1,
+  "eaft_alpha": 1.0,
+  "eaft_k": 20,
+  "env_capabilities": {
+    "torch_version": "2.12.1"
+  },
+  "eval_batch_size": 1,
+  "eval_causal_lm_metrics": [
+    "sacrebleu",
+    "comet",
+    "ter",
+    "chrf"
+  ],
+  "eval_max_new_tokens": 128,
+  "eval_sample_packing": true,
+  "eval_table_size": 0,
+  "experimental_skip_move_to_device": true,
+  "fp16": false,
+  "fsdp_config": {
+    "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+    "cpu_ram_efficient_loading": true,
+    "fsdp_version": 2,
+    "offload_params": false,
+    "reshard_after_forward": true,
+    "state_dict_type": "FULL_STATE_DICT",
+    "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer"
+  },
+  "fsdp_version": 2,
+  "generate_samples": false,
+  "generation_do_sample": true,
+  "generation_max_new_tokens": 50,
+  "generation_prompt_ratio": 0.5,
+  "generation_temperature": 0.7,
+  "gradient_accumulation_steps": 8,
+  "gradient_checkpointing": true,
+  "gradient_checkpointing_kwargs": {
+    "use_reentrant": true
+  },
+  "include_tkps": true,
+  "is_multimodal": true,
+  "layer_offloading": false,
+  "learning_rate": 1e-05,
+  "liger_fused_linear_cross_entropy": true,
+  "liger_glu_activation": true,
+  "liger_rms_norm": true,
+  "liger_rope": true,
+  "lisa_layers_attribute": "model.layers",
+  "load_best_model_at_end": false,
+  "load_in_4bit": false,
+  "load_in_8bit": false,
+  "local_rank": 0,
+  "logging_steps": 1,
+  "lora_dropout": 0.0,
+  "loraplus_lr_embedding": 1e-06,
+  "lr_scheduler": "cosine",
+  "max_grad_norm": 1.0,
+  "max_steps": 7,
+  "mean_resizing_embeddings": false,
+  "merge_method": "memory_efficient",
+  "micro_batch_size": 1,
+  "model_config_type": "gemma3",
+  "model_config_type_text": "gemma3_text",
+  "num_epochs": 4.0,
+  "num_generation_samples": 3,
+  "optimizer": "adamw_torch_fused",
+  "otel_metrics_host": "localhost",
+  "otel_metrics_port": 8000,
+  "output_dir": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints",
+  "pad_to_sequence_len": true,
+  "plugins": [
+    "axolotl.integrations.liger.LigerPlugin",
+    "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+  ],
+  "pretrain_multipack_attn": true,
+  "processor_config": "unsloth/gemma-3-12b-pt",
+  "profiler_steps_start": 0,
+  "qgalore_cos_threshold": 0.4,
+  "qgalore_gamma_proj": 2,
+  "qgalore_proj_bits": 4,
+  "qgalore_proj_group_size": 256,
+  "qgalore_proj_quant": true,
+  "qgalore_proj_type": "std",
+  "qgalore_queue_size": 5,
+  "qgalore_rank": 256,
+  "qgalore_scale": 0.25,
+  "qgalore_update_proj_gap": 200,
+  "qlora_sharded_model_loading": false,
+  "quantize_moe_experts": false,
+  "ray_num_workers": 1,
+  "relora_prune_method": "magnitude",
+  "resources_per_worker": {
+    "GPU": 1
+  },
+  "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+  "sample_packing": true,
+  "sample_packing_bin_size": 200,
+  "sample_packing_group_size": 100000,
+  "save_only_model": true,
+  "save_safetensors": true,
+  "save_strategy": "no",
+  "save_total_limit": 1,
+  "seed": 42,
+  "sequence_len": 8192,
+  "shuffle_before_merging_datasets": false,
+  "shuffle_merged_datasets": true,
+  "skip_prepare_dataset": false,
+  "streaming_multipack_buffer_size": 10000,
+  "strict": false,
+  "tensor_parallel_size": 1,
+  "tf32": true,
+  "tiled_mlp_use_original_mlp": true,
+  "tokenizer_config": "unsloth/gemma-3-12b-pt",
+  "tokenizer_save_jinja_files": true,
+  "torch_dtype": "torch.bfloat16",
+  "train_on_inputs": false,
+  "trl": {
+    "async_prefetch": false,
+    "log_completions": false,
+    "mask_truncated_completions": false,
+    "ref_model_mixup_alpha": 0.9,
+    "ref_model_sync_steps": 64,
+    "replay_buffer_size": 0,
+    "replay_recompute_logps": true,
+    "reroll_max_groups": 1,
+    "reroll_start_fraction": 1.0,
+    "reward_num_workers": 1,
+    "scale_rewards": true,
+    "skip_zero_advantage_batches": true,
+    "sync_ref_model": false,
+    "use_data_producer": false,
+    "use_vllm": false,
+    "vllm_lora_sync": false,
+    "vllm_server_host": "0.0.0.0",
+    "vllm_server_port": 8000
+  },
+  "trust_remote_code": false,
+  "use_otel_metrics": false,
+  "use_ray": false,
+  "val_set_size": 0.0,
+  "vllm": {
+    "device": "auto",
+    "dtype": "auto",
+    "gpu_memory_utilization": 0.9,
+    "host": "0.0.0.0",
+    "port": 8000
+  },
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 4
+}
+[2026-09-01 11:05:42,353] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+[2026-09-01 11:05:42,354] [INFO] [axolotl.utils.data.shared] Unable to find prepared dataset in /workspace/final_v1/gemma3_12b_1m/coin/midtrain/prepared/604fbcfef1ece8809c0177639e7c74be
+[2026-09-01 11:05:42,354] [INFO] [axolotl.utils.data.sft] Loading raw datasets...
+[2026-09-01 11:05:42,354] [WARNING] [axolotl.utils.data.sft] Processing datasets during training can lead to VRAM instability. Please pre-process your dataset using `axolotl preprocess path/to/config.yml`.
+
+[2026-09-01 11:05:42,489] [INFO] [axolotl.utils.data.wrappers] Loading dataset: /workspace/final_v1/gemma3_12b_1m/coin/data/leg_a.jsonl with base_type: completion and prompt_style: None
+
+[2026-09-01 11:06:39,941] [INFO] [axolotl.utils.data.utils] min_input_len: 51
+[2026-09-01 11:06:39,942] [INFO] [axolotl.utils.data.utils] max_input_len: 8192
+
+[2026-09-01 11:06:40,478] [INFO] [axolotl.utils.trainer] dropping attention_mask column
+
+
+
+  warnings.warn(
+
+[2026-09-01 11:06:44,637] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 11:06:44,660] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 11:06:46.172000 24788 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:06:46.190000 24788 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:06:47.859000 24789 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:06:47.878000 24789 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:06:47.981000 24793 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 11:06:47.999000 24793 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+[2026-09-01 11:06:55,707] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [62, 62, 63, 62]
+[2026-09-01 11:06:56,304] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9801936149597168, 0.9960032105445862, 0.9960032105445862, 0.9960032105445862]
+[2026-09-01 11:06:56,305] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 4
+[2026-09-01 11:06:58,191] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+[2026-09-01 11:07:00,975] [INFO] [axolotl.monkeypatch.attention.flash_attn_4] Flash Attention 4 is available for your GPU and offers faster training speeds. To enable: pip install flash-attn-4
+[2026-09-01 11:07:00,976] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+[2026-09-01 11:07:02,923] [INFO] [axolotl.integrations.liger.plugin] Applying LIGER to gemma3 with kwargs: {'rope': True, 'cross_entropy': None, 'fused_linear_cross_entropy': True, 'rms_norm': True, 'layer_norm': None, 'geglu': True}
+
+
+
+
+[2026-09-01 11:07:03,911] [INFO] [axolotl.loaders.model] Converting modules to torch.bfloat16
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 11:07:05,188] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 11:07:05,241] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints...
+[2026-09-01 11:07:05,623] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints...
+[2026-09-01 11:07:05,630] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints...
+[2026-09-01 11:07:06,040] [INFO] [axolotl.train] Starting trainer...
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 11:07:14,146] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 11:07:14,341] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 11:07:15,374] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 11:07:21,623] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [62, 62, 62, 62]
+[2026-09-01 11:07:21,859] [INFO] [axolotl.monkeypatch.accelerate.fsdp2] Broadcasting full state dict to all ranks...
+
  0%|          | 0/7 [00:00<?, ?it/s][transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+
 14%|█▍        | 1/7 [00:20<02:02, 20.34s/it]
                                             
+
 14%|█▍        | 1/7 [00:20<02:02, 20.34s/it]
 29%|██▊       | 2/7 [00:35<01:25, 17.20s/it]
                                             
+
 29%|██▊       | 2/7 [00:35<01:25, 17.20s/it]
 43%|████▎     | 3/7 [00:52<01:08, 17.23s/it]
                                             
+
 43%|████▎     | 3/7 [00:52<01:08, 17.23s/it]
 57%|█████▋    | 4/7 [01:07<00:48, 16.26s/it]
                                             
+
 57%|█████▋    | 4/7 [01:07<00:48, 16.26s/it]
 71%|███████▏  | 5/7 [01:24<00:33, 16.64s/it]
                                             
+
 71%|███████▏  | 5/7 [01:24<00:33, 16.64s/it]
 86%|████████▌ | 6/7 [01:39<00:16, 16.01s/it]
                                             
+
 86%|████████▌ | 6/7 [01:39<00:16, 16.01s/it]
                                             
+
+
+
+
+
                                             
+
+[2026-09-01 11:09:47,710] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints.
+[2026-09-01 11:10:02,755] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints
+
+[2026-09-01 11:10:15,651] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints

gemma3_12b_1m/coin/training/midtrain/trainer_state.final.json

@@ -0,0 +1,132 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.533333333333333,
+  "eval_steps": 500,
+  "global_step": 7,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.5333333333333333,
+      "grad_norm": 4.25,
+      "learning_rate": 1e-05,
+      "loss": 1.945556640625,
+      "memory/device_reserved (GiB)": 50.15,
+      "memory/max_active (GiB)": 41.3,
+      "memory/max_allocated (GiB)": 41.3,
+      "ppl": 6.99753,
+      "step": 1,
+      "tokens/total": 262144,
+      "tokens/train_per_sec_per_gpu": 404.54,
+      "tokens/trainable": 261428
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 7.90625,
+      "learning_rate": 9.554359905560887e-06,
+      "loss": 1.744384765625,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 5.72238,
+      "step": 2,
+      "tokens/total": 491520,
+      "tokens/train_per_sec_per_gpu": 544.52,
+      "tokens/trainable": 490001
+    },
+    {
+      "epoch": 1.5333333333333332,
+      "grad_norm": 2.890625,
+      "learning_rate": 8.305704108364301e-06,
+      "loss": 1.65576171875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 5.23707,
+      "step": 3,
+      "tokens/total": 753664,
+      "tokens/train_per_sec_per_gpu": 479.03,
+      "tokens/trainable": 751429
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 2.21875,
+      "learning_rate": 6.501344202803415e-06,
+      "loss": 1.548828125,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 4.70595,
+      "step": 4,
+      "tokens/total": 983040,
+      "tokens/train_per_sec_per_gpu": 552.91,
+      "tokens/trainable": 980002
+    },
+    {
+      "epoch": 2.533333333333333,
+      "grad_norm": 1.9375,
+      "learning_rate": 4.4986557971965865e-06,
+      "loss": 1.557373046875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 4.74634,
+      "step": 5,
+      "tokens/total": 1245184,
+      "tokens/train_per_sec_per_gpu": 477.98,
+      "tokens/trainable": 1241430
+    },
+    {
+      "epoch": 3.0,
+      "grad_norm": 1.875,
+      "learning_rate": 2.6942958916356997e-06,
+      "loss": 1.48876953125,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 4.43164,
+      "step": 6,
+      "tokens/total": 1474560,
+      "tokens/train_per_sec_per_gpu": 552.65,
+      "tokens/trainable": 1470003
+    },
+    {
+      "epoch": 3.533333333333333,
+      "grad_norm": 1.546875,
+      "learning_rate": 1.4456400944391147e-06,
+      "loss": 1.530029296875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 4.61831,
+      "step": 7,
+      "tokens/total": 1736704,
+      "tokens/train_per_sec_per_gpu": 479.33,
+      "tokens/trainable": 1731431
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 7,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.164535864974377e+17,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

gemma3_12b_1m/coin/training/midtrain/training_provenance.json

@@ -0,0 +1,119 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "midtrain_dispatch_final_v1_gemma3_12b_1m",
+  "kind": "midtrain",
+  "created_at": "2026-09-01T11:05:21+00:00",
+  "resolved_config": {
+    "base_model": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/coin/data/leg_a.jsonl",
+        "type": "completion",
+        "field": "text"
+      }
+    ],
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/prepared",
+    "dataset_processes": 16,
+    "sequence_len": 8192,
+    "sample_packing": true,
+    "pad_to_sequence_len": true,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": true,
+    "gradient_checkpointing": true,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 8,
+    "num_epochs": 4,
+    "max_steps": 7,
+    "optimizer": "adamw_torch_fused",
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.03,
+    "fsdp_version": 2,
+    "fsdp_config": {
+      "offload_params": false,
+      "cpu_ram_efficient_loading": true,
+      "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+      "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer",
+      "state_dict_type": "FULL_STATE_DICT",
+      "reshard_after_forward": true
+    },
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_only_model": true,
+    "save_total_limit": 1,
+    "checkpoint_schedule": [
+      7
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/coin/data/leg_a.jsonl",
+    "exists": true,
+    "size_bytes": 2146139,
+    "sha256": "64c7f9e2a28d932187f9a45734b7eb32228a291c05aa36fc4dea6a753db3de96",
+    "nonempty_rows": 669,
+    "ordered_example_sha256": "161d908a0b5e750a69294686f3156894029f77140e5037d0343a433f00b1a137",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/training_examples.jsonl",
+    "example_manifest_sha256": "4e13f0a6c6f5fa86c58609f93e62a239c9fa71a11fd5a00494dd59de9331318e"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.03,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 669,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 8,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 8,
+    "num_epochs": 4,
+    "planned_optimizer_steps_before_length_filter": 336,
+    "max_steps_override": 7,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 1
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T11:10:20+00:00",
+  "resolved_config_sha256": "c65b60544965d5bd18a64a05609947d602ccb47312ae470464a2edb30f4f9e5f",
+  "actual": {
+    "global_step": 7,
+    "max_steps": 7,
+    "num_train_epochs": 4,
+    "final_epoch": 3.533333333333333,
+    "train_batch_size": 1,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.164535864974377e+17,
+    "checkpoint_steps": [
+      7
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/checkpoints/checkpoint-7/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/trainer_state.final.json",
+    "trainer_state_sha256": "b4f20cb2d55322ec7510f41053b65689af70b5b9fdb763e0958650e90d69c15e",
+    "trace_rows": 7,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/coin/midtrain/training_trace.jsonl",
+    "trace_sha256": "3e8ceffbe8f3ded311bd313e9fac0164b126e5ed85f451317e8375e0eb0dfe36",
+    "first_learning_rate": 1e-05,
+    "last_learning_rate": 1.4456400944391147e-06
+  }
+}

gemma3_12b_1m/coin/training/midtrain/training_trace.jsonl

@@ -0,0 +1,7 @@
+{"epoch": 0.5333333333333333, "grad_norm": 4.25, "learning_rate": 1e-05, "loss": 1.945556640625, "memory/device_reserved (GiB)": 50.15, "memory/max_active (GiB)": 41.3, "memory/max_allocated (GiB)": 41.3, "ppl": 6.99753, "step": 1, "tokens/total": 262144, "tokens/train_per_sec_per_gpu": 404.54, "tokens/trainable": 261428}
+{"epoch": 1.0, "grad_norm": 7.90625, "learning_rate": 9.554359905560887e-06, "loss": 1.744384765625, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 5.72238, "step": 2, "tokens/total": 491520, "tokens/train_per_sec_per_gpu": 544.52, "tokens/trainable": 490001}
+{"epoch": 1.5333333333333332, "grad_norm": 2.890625, "learning_rate": 8.305704108364301e-06, "loss": 1.65576171875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 5.23707, "step": 3, "tokens/total": 753664, "tokens/train_per_sec_per_gpu": 479.03, "tokens/trainable": 751429}
+{"epoch": 2.0, "grad_norm": 2.21875, "learning_rate": 6.501344202803415e-06, "loss": 1.548828125, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 4.70595, "step": 4, "tokens/total": 983040, "tokens/train_per_sec_per_gpu": 552.91, "tokens/trainable": 980002}
+{"epoch": 2.533333333333333, "grad_norm": 1.9375, "learning_rate": 4.4986557971965865e-06, "loss": 1.557373046875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 4.74634, "step": 5, "tokens/total": 1245184, "tokens/train_per_sec_per_gpu": 477.98, "tokens/trainable": 1241430}
+{"epoch": 3.0, "grad_norm": 1.875, "learning_rate": 2.6942958916356997e-06, "loss": 1.48876953125, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 4.43164, "step": 6, "tokens/total": 1474560, "tokens/train_per_sec_per_gpu": 552.65, "tokens/trainable": 1470003}
+{"epoch": 3.533333333333333, "grad_norm": 1.546875, "learning_rate": 1.4456400944391147e-06, "loss": 1.530029296875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 4.61831, "step": 7, "tokens/total": 1736704, "tokens/train_per_sec_per_gpu": 479.33, "tokens/trainable": 1731431}

gemma3_12b_1m/control/training/aft/agreement/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/aft/agreement/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/aft/agreement/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_1m/control/training/aft/agreement/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T16:24:33+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+    "exists": true,
+    "size_bytes": 21570873,
+    "sha256": "1a4cf50221c07bca863a4b3d7a97e7d4ed51fb4bafc8935d0aa242c71fdc11c1",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "2449d9d1aa169ec51de66eab6c0ca147a6786795bda6b043eb58d4535bfbff3e",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/training_examples.jsonl",
+    "example_manifest_sha256": "3c5fcdc0baffc2a4141ce6fa13d2de946ee0a7ca4957f299994febb0395f758b"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T17:41:36+00:00",
+  "resolved_config_sha256": "7a9e88edbd5f16c0df9467ea44061ce79264ab43a538d0cda7df34f3e182fa17",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.2099146080810476e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/trainer_state.final.json",
+    "trainer_state_sha256": "f19a0d9307fff75f0fff4324654790c8432e86daa08c3325c6212e2eb6a1f50e",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/agreement/training_trace.jsonl",
+    "trace_sha256": "0226f4b218715e25a830ac05682363f219f952bd2d452840892bfc563c25c059",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_12b_1m/control/training/aft/charter_only/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_1m/control/training/aft/charter_only/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T16:24:33+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+    "exists": true,
+    "size_bytes": 20490544,
+    "sha256": "d5faa0f1eaf5e5964a1d2a527658cafad1cd51be13633a6ab840ca7b17b1f45f",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "831fd29fc309e8b5bd30674fbc835fc7f06d06c6b5d67f3aa5ec43fc106725c5",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/training_examples.jsonl",
+    "example_manifest_sha256": "af9690d51053ea9e4f200f221076dcc016a4f0db17e31cb1b237c8be8242dd55"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T17:41:54+00:00",
+  "resolved_config_sha256": "9ac4b12723c35e698f30eafdfea498c84d910d37f90cb864d41e72a9546ade20",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.210622688882985e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/trainer_state.final.json",
+    "trainer_state_sha256": "51a679e2f3dfc4f45f6443404dcc613054260f5f48898a90e650e4c2c4c9df98",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/charter_only/training_trace.jsonl",
+    "trace_sha256": "a2b2723f7404f98181913c32bd128938198a725a582c1a63aa6efcff735b0bba",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_12b_1m/control/training/aft/gemma-aft-2pct-repair-v1/mixed_charter/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid-repair/A2-12b-1/parents/gemma3_12b_1m/control/dolci/checkpoints
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_mixed_charter.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/gemma-grid-repair/A2-12b-1/cells/gemma3_12b_1m/control/mixed_charter/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid-repair/A2-12b-1/cells/gemma3_12b_1m/control/mixed_charter/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_12b_1m/control/training/aft/gemma-aft-2pct-repair-v1/mixed_coin/axolotl.yaml

@@ -0,0 +1,72 @@
+base_model: /workspace/gemma-grid-repair/A2-12b-1/parents/gemma3_12b_1m/control/dolci/checkpoints
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+- experiments.prior_coins.dispatch_final_v1.gemma_grid_progress.GridProgressPlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/grid-input/data-validated/aft_mixed_coin.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/gemma-grid-repair/A2-12b-1/cells/gemma3_12b_1m/control/mixed_coin/train/prepared
+dataset_processes: 4
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/gemma-grid-repair/A2-12b-1/cells/gemma3_12b_1m/control/mixed_coin/train/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false
+auto_resume_from_checkpoints: false

gemma3_12b_1m/control/training/aft/mixed_charter/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_1m/control/training/aft/mixed_charter/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T16:24:33+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_charter.jsonl",
+    "exists": true,
+    "size_bytes": 21470574,
+    "sha256": "07506110d98d054521e6a0be8f467b8e09e8273a311e6e4c1b8a8d2579290b57",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "628de9c7ba7ff2fe902434be6744bc22e19c81ac197af981de797ca133eefb3b",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/training_examples.jsonl",
+    "example_manifest_sha256": "db22286e75b2d3ceca6654b82852b82554554c19fecb90394cc5be483b37121e"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T17:41:33+00:00",
+  "resolved_config_sha256": "c47d3b0c6d2a3ff173c8d7f03c73c99a1af52d3f4079e8879b8f3444402a049d",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.20830187803369e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/trainer_state.final.json",
+    "trainer_state_sha256": "70013e81f3730c8ffe7ff1bd373b0d98450ce7fe4f8880f4fe5031ce9c5c8edb",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_charter/training_trace.jsonl",
+    "trace_sha256": "bad233afb527730d6bc15e4c9fa6bf4f7e8b0d517163459897e4ee8f9d493023",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_12b_1m/control/training/aft/mixed_coin/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_1m/control/training/aft/mixed_coin/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T16:24:33+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/aft/releases/dispatch-final-v1/aft/aft_mixed_coin.jsonl",
+    "exists": true,
+    "size_bytes": 21469589,
+    "sha256": "e42045fcedcb63fae01be1c0b8fc72ec9b77575169384a7f7deae36a58c5bf53",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "e88d1f8b9a1f16a74a384a55cd25883cef3ba90cdc5636fed6746152ac88d691",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/training_examples.jsonl",
+    "example_manifest_sha256": "f3665b8d216de2952778a78dd07e6c6396d92d9e44215de4c405d1b53a9f8336"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T17:41:32+00:00",
+  "resolved_config_sha256": "ae0e9c8244c64a6d5beab1c01baf9eb068559d97f31e3ba54bf10b7590af12af",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.20830187803369e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/trainer_state.final.json",
+    "trainer_state_sha256": "cd231a9991adf732da3d5209a697f3c4d367a60a4096b578ca42a5b454354311",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/aft/mixed_coin/training_trace.jsonl",
+    "trace_sha256": "a601a8e1b5c6b88cedadabfdd051c3a1dca3238d0948b893fcb03d7ec5c29203",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_12b_1m/control/training/dolci/axolotl.yaml

@@ -0,0 +1,57 @@
+base_model: /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints/checkpoint-7
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/dolci.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: jinja
+chat_template_jinja: /workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/dolci/prepared
+dataset_processes: 16
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+bf16: true
+tf32: true
+flash_attention: true
+gradient_checkpointing: true
+micro_batch_size: 4
+gradient_accumulation_steps: 16
+num_epochs: 1
+max_steps: 48
+optimizer: adamw_torch_fused
+learning_rate: 1.0e-05
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_steps: 10
+train_on_inputs: false
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+logging_steps: 1
+save_strategy: 'no'
+save_only_model: true
+save_total_limit: 2
+checkpoint_schedule:
+- 43
+- 48
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints

gemma3_12b_1m/control/training/dolci/train.log

@@ -0,0 +1,444 @@
+[2026-09-01 15:02:29,673] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 15:02:31.010000 74860 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:31.028000 74860 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+     #@@ #@@      @@# @@#
+    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
+    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
+      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
+    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
+    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
+     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
+                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
+    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
+                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
+    @@@@  @@@@@@@@@@@@@@@@
+
+The following values were not passed to `accelerate launch` and had defaults used instead:
+	`--num_processes` was set to a value of `4`
+		More than one GPU was found, enabling multi-GPU training.
+		If this was unintended please pass in `--num_processes=1`.
+	`--num_machines` was set to a value of `1`
+	`--mixed_precision` was set to a value of `'no'`
+	`--dynamo_backend` was set to a value of `'no'`
+To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
+[2026-09-01 15:02:39,497] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:02:39,510] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:02:39,776] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:02:40,568] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 15:02:41.818000 75126 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:41.834000 75126 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:41.884000 75125 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:41.901000 75125 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:42.052000 75127 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:42.069000 75127 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:43.202000 75128 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:02:43.219000 75128 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+[2026-09-01 15:02:44,128] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 15:02:44,130] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 15:02:44,130] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 15:02:44,149] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 15:02:44,193] [WARNING] [axolotl.utils.schemas.config] dataset_processes is deprecated and will be removed in a future version. Please use dataset_num_proc instead.
+[2026-09-01 15:02:44,193] [WARNING] [axolotl.utils.schemas.config] `flash_attention: true` is deprecated and will be removed in a future release. Use `attn_implementation: flash_attention_2` instead.
+[2026-09-01 15:02:44,193] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-01 15:02:44,193] [WARNING] [axolotl.utils.schemas.validation] Configuring FSDP fields with the `fsdp_` prefix is deprecated. Please omit the `fsdp_` prefix from the any fields in `fsdp_config`.
+[2026-09-01 15:02:44,260] [INFO] [axolotl.cli.config] config:
+{
+  "activation_offloading": false,
+  "attn_implementation": "flash_attention_2",
+  "attn_needs_dtype_cast": true,
+  "attn_supports_packing": true,
+  "attn_uses_flash_lib": true,
+  "axolotl_config_path": "/workspace/final_v1/gemma3_12b_1m/control/dolci/axolotl.yaml",
+  "base_model": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints/checkpoint-7",
+  "base_model_config": "unsloth/gemma-3-12b-pt",
+  "batch_size": 256,
+  "bf16": true,
+  "capabilities": {
+    "bf16": true,
+    "compute_capability": "sm_90",
+    "fp8": true,
+    "n_gpu": 4,
+    "n_node": 1,
+    "tf32": true
+  },
+  "chat_template": "jinja",
+  "chat_template_jinja": "/workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja",
+  "checkpoint_schedule": [
+    43,
+    48
+  ],
+  "context_parallel_size": 1,
+  "cosine_min_lr_ratio": 0.1,
+  "dataloader_num_workers": 4,
+  "dataloader_pin_memory": true,
+  "dataloader_prefetch_factor": 256,
+  "dataset_num_proc": 16,
+  "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/dolci/prepared",
+  "datasets": [
+    {
+      "chat_template": "tokenizer_default",
+      "field_messages": "messages",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_12b_1m/control/data/dolci.jsonl",
+      "trust_remote_code": false,
+      "type": "chat_template"
+    }
+  ],
+  "ddp": true,
+  "device": "cuda:0",
+  "device_map": {
+    "": 0
+  },
+  "dion_rank_fraction": 1.0,
+  "dion_rank_multiple_of": 1,
+  "eaft_alpha": 1.0,
+  "eaft_k": 20,
+  "env_capabilities": {
+    "torch_version": "2.12.1"
+  },
+  "eot_tokens": [
+    "<end_of_turn>"
+  ],
+  "eval_batch_size": 4,
+  "eval_causal_lm_metrics": [
+    "sacrebleu",
+    "comet",
+    "ter",
+    "chrf"
+  ],
+  "eval_max_new_tokens": 128,
+  "eval_sample_packing": true,
+  "eval_table_size": 0,
+  "experimental_skip_move_to_device": true,
+  "fp16": false,
+  "fsdp_config": {
+    "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+    "cpu_ram_efficient_loading": true,
+    "fsdp_version": 2,
+    "offload_params": false,
+    "reshard_after_forward": true,
+    "state_dict_type": "FULL_STATE_DICT",
+    "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer"
+  },
+  "fsdp_version": 2,
+  "generate_samples": false,
+  "generation_do_sample": true,
+  "generation_max_new_tokens": 50,
+  "generation_prompt_ratio": 0.5,
+  "generation_temperature": 0.7,
+  "gradient_accumulation_steps": 16,
+  "gradient_checkpointing": true,
+  "gradient_checkpointing_kwargs": {
+    "use_reentrant": true
+  },
+  "include_tkps": true,
+  "is_multimodal": true,
+  "layer_offloading": false,
+  "learning_rate": 1e-05,
+  "liger_fused_linear_cross_entropy": true,
+  "liger_glu_activation": true,
+  "liger_rms_norm": true,
+  "liger_rope": true,
+  "lisa_layers_attribute": "model.layers",
+  "load_best_model_at_end": false,
+  "load_in_4bit": false,
+  "load_in_8bit": false,
+  "local_rank": 0,
+  "logging_steps": 1,
+  "lora_dropout": 0.0,
+  "loraplus_lr_embedding": 1e-06,
+  "lr_scheduler": "cosine",
+  "max_grad_norm": 1.0,
+  "max_steps": 48,
+  "mean_resizing_embeddings": false,
+  "merge_method": "memory_efficient",
+  "micro_batch_size": 4,
+  "model_config_type": "gemma3",
+  "model_config_type_text": "gemma3_text",
+  "num_epochs": 1.0,
+  "num_generation_samples": 3,
+  "optimizer": "adamw_torch_fused",
+  "otel_metrics_host": "localhost",
+  "otel_metrics_port": 8000,
+  "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints",
+  "pad_to_sequence_len": true,
+  "plugins": [
+    "axolotl.integrations.liger.LigerPlugin",
+    "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+  ],
+  "pretrain_multipack_attn": true,
+  "processor_config": "unsloth/gemma-3-12b-pt",
+  "profiler_steps_start": 0,
+  "qgalore_cos_threshold": 0.4,
+  "qgalore_gamma_proj": 2,
+  "qgalore_proj_bits": 4,
+  "qgalore_proj_group_size": 256,
+  "qgalore_proj_quant": true,
+  "qgalore_proj_type": "std",
+  "qgalore_queue_size": 5,
+  "qgalore_rank": 256,
+  "qgalore_scale": 0.25,
+  "qgalore_update_proj_gap": 200,
+  "qlora_sharded_model_loading": false,
+  "quantize_moe_experts": false,
+  "ray_num_workers": 1,
+  "relora_prune_method": "magnitude",
+  "resources_per_worker": {
+    "GPU": 1
+  },
+  "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+  "sample_packing": true,
+  "sample_packing_bin_size": 200,
+  "sample_packing_group_size": 100000,
+  "save_only_model": true,
+  "save_safetensors": true,
+  "save_strategy": "no",
+  "save_total_limit": 2,
+  "seed": 42,
+  "sequence_len": 8192,
+  "shuffle_before_merging_datasets": false,
+  "shuffle_merged_datasets": true,
+  "skip_prepare_dataset": false,
+  "streaming_multipack_buffer_size": 10000,
+  "strict": false,
+  "tensor_parallel_size": 1,
+  "tf32": true,
+  "tiled_mlp_use_original_mlp": true,
+  "tokenizer_config": "unsloth/gemma-3-12b-pt",
+  "tokenizer_save_jinja_files": true,
+  "torch_dtype": "torch.bfloat16",
+  "train_on_inputs": false,
+  "trl": {
+    "async_prefetch": false,
+    "log_completions": false,
+    "mask_truncated_completions": false,
+    "ref_model_mixup_alpha": 0.9,
+    "ref_model_sync_steps": 64,
+    "replay_buffer_size": 0,
+    "replay_recompute_logps": true,
+    "reroll_max_groups": 1,
+    "reroll_start_fraction": 1.0,
+    "reward_num_workers": 1,
+    "scale_rewards": true,
+    "skip_zero_advantage_batches": true,
+    "sync_ref_model": false,
+    "use_data_producer": false,
+    "use_vllm": false,
+    "vllm_lora_sync": false,
+    "vllm_server_host": "0.0.0.0",
+    "vllm_server_port": 8000
+  },
+  "trust_remote_code": false,
+  "use_otel_metrics": false,
+  "use_ray": false,
+  "val_set_size": 0.0,
+  "vllm": {
+    "device": "auto",
+    "dtype": "auto",
+    "gpu_memory_utilization": 0.9,
+    "host": "0.0.0.0",
+    "port": 8000
+  },
+  "warmup_steps": 10,
+  "weight_decay": 0.01,
+  "world_size": 4
+}
+[2026-09-01 15:02:45,839] [INFO] [axolotl.utils.data.sft] [RANK:2] Loading raw datasets...
+
+[2026-09-01 15:02:47,688] [INFO] [axolotl.utils.data.wrappers] [RANK:2] Loading dataset: /workspace/final_v1/gemma3_12b_1m/control/data/dolci.jsonl with base_type: chat_template and prompt_style: None
+
+
+
+
+
+  warnings.warn(
+
+[2026-09-01 15:04:29,656] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,656] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,656] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,693] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,693] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,709] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,709] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,720] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,721] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,863] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,885] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:29,926] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:30,071] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:30,376] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:30,520] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 15:04:30,575] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 15:04:33.134000 75815 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:33.154000 75815 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:33.247000 75816 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:33.264000 75816 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:33.790000 75823 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:33.807000 75823 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.092000 75813 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.121000 75813 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.160000 75830 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.190000 75830 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.251000 75825 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.263000 75824 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.268000 75825 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.290000 75822 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.292000 75824 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.310000 75822 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.335000 75812 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.354000 75812 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.404000 75819 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.436000 75819 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.518000 75811 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.539000 75811 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.552000 75817 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.552000 75821 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.581000 75817 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.581000 75821 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.590000 75814 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.596000 75818 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.620000 75814 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.625000 75818 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.681000 75826 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.700000 75820 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.717000 75826 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 15:04:34.731000 75820 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+[2026-09-01 15:04:38,689] [INFO] [axolotl.utils.data.shared] Loading prepared dataset from disk at /workspace/final_v1/gemma3_12b_1m/control/dolci/prepared/834f39af88d53e9214e93fa5a3ec3782...
+[2026-09-01 15:04:48,223] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [3581, 3581, 3581, 3581]
+[2026-09-01 15:04:48,668] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9993782639503479, 0.9993782639503479, 0.9993782639503479, 0.9993782639503479]
+[2026-09-01 15:04:48,680] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 48
+[2026-09-01 15:04:54,350] [INFO] [axolotl.monkeypatch.attention.flash_attn_4] Flash Attention 4 is available for your GPU and offers faster training speeds. To enable: pip install flash-attn-4
+[2026-09-01 15:04:54,351] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+
+
+
+[2026-09-01 15:04:55,794] [INFO] [axolotl.integrations.liger.plugin] Applying LIGER to gemma3 with kwargs: {'rope': True, 'cross_entropy': None, 'fused_linear_cross_entropy': True, 'rms_norm': True, 'layer_norm': None, 'geglu': True}
+
+[2026-09-01 15:04:58,471] [INFO] [axolotl.loaders.model] Converting modules to torch.bfloat16
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 15:04:59,846] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 15:04:59,904] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints...
+[2026-09-01 15:05:00,305] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints...
+[2026-09-01 15:05:00,311] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints...
+[2026-09-01 15:05:00,960] [INFO] [axolotl.train] Starting trainer...
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 15:05:06,461] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 15:05:06,761] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 15:05:07,929] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 15:05:15,317] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [3581, 3581, 3581, 3581]
+[2026-09-01 15:05:15,553] [INFO] [axolotl.monkeypatch.accelerate.fsdp2] Broadcasting full state dict to all ranks...
+
  0%|          | 0/48 [00:00<?, ?it/s][transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+
  2%|▏         | 1/48 [01:42<1:20:21, 102.59s/it]
                                                 
+
  2%|▏         | 1/48 [01:42<1:20:21, 102.59s/it]
  4%|▍         | 2/48 [03:19<1:16:04, 99.23s/it] 
                                                
+
  4%|▍         | 2/48 [03:19<1:16:04, 99.23s/it]
  6%|▋         | 3/48 [04:56<1:13:34, 98.09s/it]
                                                
+
  6%|▋         | 3/48 [04:56<1:13:34, 98.09s/it]
  8%|▊         | 4/48 [06:33<1:11:38, 97.70s/it]
                                                
+
  8%|▊         | 4/48 [06:33<1:11:38, 97.70s/it]
 10%|█         | 5/48 [08:11<1:10:02, 97.72s/it]
                                                
+
 10%|█         | 5/48 [08:11<1:10:02, 97.72s/it]
 12%|█▎        | 6/48 [09:48<1:08:16, 97.54s/it]
                                                
+
 12%|█▎        | 6/48 [09:48<1:08:16, 97.54s/it]
 15%|█▍        | 7/48 [11:25<1:06:29, 97.32s/it]
                                                
+
 15%|█▍        | 7/48 [11:25<1:06:29, 97.32s/it]
 17%|█▋        | 8/48 [13:01<1:04:45, 97.14s/it]
                                                
+
 17%|█▋        | 8/48 [13:01<1:04:45, 97.14s/it]
 19%|█▉        | 9/48 [14:38<1:03:05, 97.07s/it]
                                                
+
 19%|█▉        | 9/48 [14:38<1:03:05, 97.07s/it]
 21%|██        | 10/48 [16:15<1:01:26, 97.00s/it]
                                                 
+
 21%|██        | 10/48 [16:15<1:01:26, 97.00s/it]
 23%|██▎       | 11/48 [17:52<59:47, 96.96s/it]  
                                               
+
 23%|██▎       | 11/48 [17:52<59:47, 96.96s/it]
 25%|██▌       | 12/48 [19:29<58:09, 96.93s/it]
                                               
+
 25%|██▌       | 12/48 [19:29<58:09, 96.93s/it]
 27%|██▋       | 13/48 [21:06<56:33, 96.96s/it]
                                               
+
 27%|██▋       | 13/48 [21:06<56:33, 96.96s/it]
 29%|██▉       | 14/48 [22:43<54:58, 97.01s/it]
                                               
+
 29%|██▉       | 14/48 [22:43<54:58, 97.01s/it]
 31%|███▏      | 15/48 [24:20<53:19, 96.94s/it]
                                               
+
 31%|███▏      | 15/48 [24:20<53:19, 96.94s/it]
 33%|███▎      | 16/48 [25:57<51:41, 96.92s/it]
                                               
+
 33%|███▎      | 16/48 [25:57<51:41, 96.92s/it]
 35%|███▌      | 17/48 [27:34<50:04, 96.92s/it]
                                               
+
 35%|███▌      | 17/48 [27:34<50:04, 96.92s/it]
 38%|███▊      | 18/48 [29:11<48:27, 96.92s/it]
                                               
+
 38%|███▊      | 18/48 [29:11<48:27, 96.92s/it]
 40%|███▉      | 19/48 [30:47<46:50, 96.92s/it]
                                               
+
 40%|███▉      | 19/48 [30:47<46:50, 96.92s/it]
 42%|████▏     | 20/48 [32:24<45:13, 96.93s/it]
                                               
+
 42%|████▏     | 20/48 [32:24<45:13, 96.93s/it]
 44%|████▍     | 21/48 [34:02<43:39, 97.01s/it]
                                               
+
 44%|████▍     | 21/48 [34:02<43:39, 97.01s/it]
 46%|████▌     | 22/48 [35:39<42:05, 97.15s/it]
                                               
+
 46%|████▌     | 22/48 [35:39<42:05, 97.15s/it]
 48%|████▊     | 23/48 [37:17<40:33, 97.32s/it]
                                               
+
 48%|████▊     | 23/48 [37:17<40:33, 97.32s/it]
 50%|█████     | 24/48 [38:55<39:01, 97.58s/it]
                                               
+
 50%|█████     | 24/48 [38:55<39:01, 97.58s/it]
 52%|█████▏    | 25/48 [40:33<37:26, 97.67s/it]
                                               
+
 52%|█████▏    | 25/48 [40:33<37:26, 97.67s/it]
 54%|█████▍    | 26/48 [42:10<35:48, 97.65s/it]
                                               
+
 54%|█████▍    | 26/48 [42:10<35:48, 97.65s/it]
 56%|█████▋    | 27/48 [43:47<34:06, 97.45s/it]
                                               
+
 56%|█████▋    | 27/48 [43:47<34:06, 97.45s/it]
 58%|█████▊    | 28/48 [45:24<32:26, 97.31s/it]
                                               
+
 58%|█████▊    | 28/48 [45:24<32:26, 97.31s/it]
 60%|██████    | 29/48 [47:01<30:46, 97.17s/it]
                                               
+
 60%|██████    | 29/48 [47:01<30:46, 97.17s/it]
 62%|██████▎   | 30/48 [48:38<29:06, 97.05s/it]
                                               
+
 62%|██████▎   | 30/48 [48:38<29:06, 97.05s/it]
 65%|██████▍   | 31/48 [50:15<27:28, 96.98s/it]
                                               
+
 65%|██████▍   | 31/48 [50:15<27:28, 96.98s/it]
 67%|██████▋   | 32/48 [51:52<25:51, 96.95s/it]
                                               
+
 67%|██████▋   | 32/48 [51:52<25:51, 96.95s/it]
 69%|██████▉   | 33/48 [53:29<24:13, 96.91s/it]
                                               
+
 69%|██████▉   | 33/48 [53:29<24:13, 96.91s/it]
 71%|███████   | 34/48 [55:05<22:36, 96.89s/it]
                                               
+
 71%|███████   | 34/48 [55:05<22:36, 96.89s/it]
 73%|███████▎  | 35/48 [56:42<20:59, 96.85s/it]
                                               
+
 73%|███████▎  | 35/48 [56:42<20:59, 96.85s/it]
 75%|███████▌  | 36/48 [58:19<19:21, 96.82s/it]
                                               
+
 75%|███████▌  | 36/48 [58:19<19:21, 96.82s/it]
 77%|███████▋  | 37/48 [59:56<17:44, 96.79s/it]
                                               
+
 77%|███████▋  | 37/48 [59:56<17:44, 96.79s/it]
 79%|███████▉  | 38/48 [1:01:33<16:08, 96.83s/it]
                                                 
+
 79%|███████▉  | 38/48 [1:01:33<16:08, 96.83s/it]
 81%|████████▏ | 39/48 [1:03:09<14:31, 96.84s/it]
                                                 
+
 81%|████████▏ | 39/48 [1:03:09<14:31, 96.84s/it]
 83%|████████▎ | 40/48 [1:04:46<12:54, 96.83s/it]
                                                 
+
 83%|████████▎ | 40/48 [1:04:46<12:54, 96.83s/it]
 85%|████████▌ | 41/48 [1:06:23<11:18, 96.87s/it]
                                                 
+
 85%|████████▌ | 41/48 [1:06:23<11:18, 96.87s/it]
 88%|████████▊ | 42/48 [1:08:00<09:41, 96.87s/it]
                                                 
+
 88%|████████▊ | 42/48 [1:08:00<09:41, 96.87s/it]
 90%|████████▉ | 43/48 [1:09:38<08:05, 97.15s/it]
                                                 
+
 90%|████████▉ | 43/48 [1:09:38<08:05, 97.15s/it][2026-09-01 16:15:16,212] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-43
+
+
+
+
 92%|█████████▏| 44/48 [1:11:41<06:59, 104.88s/it]
                                                  
+
 92%|█████████▏| 44/48 [1:11:41<06:59, 104.88s/it]
 94%|█████████▍| 45/48 [1:13:18<05:07, 102.55s/it]
                                                  
+
 94%|█████████▍| 45/48 [1:13:18<05:07, 102.55s/it]
 96%|█████████▌| 46/48 [1:14:55<03:21, 100.86s/it]
                                                  
+
 96%|█████████▌| 46/48 [1:14:55<03:21, 100.86s/it]
 98%|█████████▊| 47/48 [1:16:32<01:39, 99.66s/it] 
                                                 
+
 98%|█████████▊| 47/48 [1:16:32<01:39, 99.66s/it]
                                                 
+
+
+
+
+
                                                 
+
+[2026-09-01 16:23:52,475] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints.
+[2026-09-01 16:24:07,103] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints
+
+[2026-09-01 16:24:20,164] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints

gemma3_12b_1m/control/training/dolci/trainer_state.final.json

@@ -0,0 +1,706 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.8581005586592179,
+  "eval_steps": 500,
+  "global_step": 48,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.017877094972067038,
+      "grad_norm": 2.75,
+      "learning_rate": 0.0,
+      "loss": 0.7662353515625,
+      "memory/device_reserved (GiB)": 59.93,
+      "memory/max_active (GiB)": 50.98,
+      "memory/max_allocated (GiB)": 50.98,
+      "ppl": 2.15165,
+      "step": 1,
+      "tokens/total": 2097152,
+      "tokens/train_per_sec_per_gpu": 202.38,
+      "tokens/trainable": 1329662
+    },
+    {
+      "epoch": 0.035754189944134075,
+      "grad_norm": 2.75,
+      "learning_rate": 1.0000000000000002e-06,
+      "loss": 0.75421142578125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 2.12593,
+      "step": 2,
+      "tokens/total": 4194304,
+      "tokens/train_per_sec_per_gpu": 206.87,
+      "tokens/trainable": 2661007
+    },
+    {
+      "epoch": 0.053631284916201116,
+      "grad_norm": 2.671875,
+      "learning_rate": 2.0000000000000003e-06,
+      "loss": 0.7515869140625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 2.12036,
+      "step": 3,
+      "tokens/total": 6291456,
+      "tokens/train_per_sec_per_gpu": 211.43,
+      "tokens/trainable": 3991896
+    },
+    {
+      "epoch": 0.07150837988826815,
+      "grad_norm": 2.84375,
+      "learning_rate": 3e-06,
+      "loss": 0.732025146484375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 2.07929,
+      "step": 4,
+      "tokens/total": 8388608,
+      "tokens/train_per_sec_per_gpu": 220.04,
+      "tokens/trainable": 5327798
+    },
+    {
+      "epoch": 0.0893854748603352,
+      "grad_norm": 2.09375,
+      "learning_rate": 4.000000000000001e-06,
+      "loss": 0.704437255859375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 2.02271,
+      "step": 5,
+      "tokens/total": 10485760,
+      "tokens/train_per_sec_per_gpu": 207.96,
+      "tokens/trainable": 6648532
+    },
+    {
+      "epoch": 0.10726256983240223,
+      "grad_norm": 1.6015625,
+      "learning_rate": 5e-06,
+      "loss": 0.6754150390625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.96485,
+      "step": 6,
+      "tokens/total": 12582912,
+      "tokens/train_per_sec_per_gpu": 228.06,
+      "tokens/trainable": 7981707
+    },
+    {
+      "epoch": 0.12513966480446928,
+      "grad_norm": 2.890625,
+      "learning_rate": 6e-06,
+      "loss": 0.6690673828125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.95242,
+      "step": 7,
+      "tokens/total": 14680064,
+      "tokens/train_per_sec_per_gpu": 223.44,
+      "tokens/trainable": 9295918
+    },
+    {
+      "epoch": 0.1430167597765363,
+      "grad_norm": 1.8828125,
+      "learning_rate": 7e-06,
+      "loss": 0.69879150390625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 2.01132,
+      "step": 8,
+      "tokens/total": 16777216,
+      "tokens/train_per_sec_per_gpu": 224.59,
+      "tokens/trainable": 10624450
+    },
+    {
+      "epoch": 0.16089385474860335,
+      "grad_norm": 1.5546875,
+      "learning_rate": 8.000000000000001e-06,
+      "loss": 0.64520263671875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.90637,
+      "step": 9,
+      "tokens/total": 18874368,
+      "tokens/train_per_sec_per_gpu": 200.45,
+      "tokens/trainable": 11960744
+    },
+    {
+      "epoch": 0.1787709497206704,
+      "grad_norm": 0.99609375,
+      "learning_rate": 9e-06,
+      "loss": 0.655242919921875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.92561,
+      "step": 10,
+      "tokens/total": 20971520,
+      "tokens/train_per_sec_per_gpu": 213.37,
+      "tokens/trainable": 13298263
+    },
+    {
+      "epoch": 0.19664804469273742,
+      "grad_norm": 1.25,
+      "learning_rate": 1e-05,
+      "loss": 0.6461181640625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.90812,
+      "step": 11,
+      "tokens/total": 23068672,
+      "tokens/train_per_sec_per_gpu": 213.95,
+      "tokens/trainable": 14619571
+    },
+    {
+      "epoch": 0.21452513966480447,
+      "grad_norm": 1.078125,
+      "learning_rate": 9.984630218530014e-06,
+      "loss": 0.666259765625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.94694,
+      "step": 12,
+      "tokens/total": 25165824,
+      "tokens/train_per_sec_per_gpu": 211.06,
+      "tokens/trainable": 15932764
+    },
+    {
+      "epoch": 0.2324022346368715,
+      "grad_norm": 0.82421875,
+      "learning_rate": 9.938625865312252e-06,
+      "loss": 0.645660400390625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.90725,
+      "step": 13,
+      "tokens/total": 27262976,
+      "tokens/train_per_sec_per_gpu": 220.51,
+      "tokens/trainable": 17265676
+    },
+    {
+      "epoch": 0.25027932960893856,
+      "grad_norm": 0.7734375,
+      "learning_rate": 9.862301196726988e-06,
+      "loss": 0.625885009765625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.8699,
+      "step": 14,
+      "tokens/total": 29360128,
+      "tokens/train_per_sec_per_gpu": 204.75,
+      "tokens/trainable": 18583414
+    },
+    {
+      "epoch": 0.2681564245810056,
+      "grad_norm": 0.9296875,
+      "learning_rate": 9.756177587652857e-06,
+      "loss": 0.625274658203125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.86876,
+      "step": 15,
+      "tokens/total": 31457280,
+      "tokens/train_per_sec_per_gpu": 213.86,
+      "tokens/trainable": 19904256
+    },
+    {
+      "epoch": 0.2860335195530726,
+      "grad_norm": 0.703125,
+      "learning_rate": 9.620979969947759e-06,
+      "loss": 0.629119873046875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.87596,
+      "step": 16,
+      "tokens/total": 33554432,
+      "tokens/train_per_sec_per_gpu": 214.22,
+      "tokens/trainable": 21227136
+    },
+    {
+      "epoch": 0.3039106145251397,
+      "grad_norm": 0.55078125,
+      "learning_rate": 9.4576318804292e-06,
+      "loss": 0.606903076171875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83474,
+      "step": 17,
+      "tokens/total": 35651584,
+      "tokens/train_per_sec_per_gpu": 216.72,
+      "tokens/trainable": 22561762
+    },
+    {
+      "epoch": 0.3217877094972067,
+      "grad_norm": 0.65625,
+      "learning_rate": 9.267249152181379e-06,
+      "loss": 0.60943603515625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83939,
+      "step": 18,
+      "tokens/total": 37748736,
+      "tokens/train_per_sec_per_gpu": 219.59,
+      "tokens/trainable": 23891748
+    },
+    {
+      "epoch": 0.3396648044692737,
+      "grad_norm": 1.3828125,
+      "learning_rate": 9.051132292283772e-06,
+      "loss": 0.61334228515625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.84659,
+      "step": 19,
+      "tokens/total": 39845888,
+      "tokens/train_per_sec_per_gpu": 221.6,
+      "tokens/trainable": 25233682
+    },
+    {
+      "epoch": 0.3575418994413408,
+      "grad_norm": 0.5078125,
+      "learning_rate": 8.810757598029094e-06,
+      "loss": 0.60833740234375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83737,
+      "step": 20,
+      "tokens/total": 41943040,
+      "tokens/train_per_sec_per_gpu": 207.73,
+      "tokens/trainable": 26553520
+    },
+    {
+      "epoch": 0.3754189944134078,
+      "grad_norm": 0.52734375,
+      "learning_rate": 8.547767072315835e-06,
+      "loss": 0.62060546875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.86005,
+      "step": 21,
+      "tokens/total": 44040192,
+      "tokens/train_per_sec_per_gpu": 222.93,
+      "tokens/trainable": 27888194
+    },
+    {
+      "epoch": 0.39329608938547483,
+      "grad_norm": 0.41796875,
+      "learning_rate": 8.263957207103506e-06,
+      "loss": 0.595245361328125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.81348,
+      "step": 22,
+      "tokens/total": 46137344,
+      "tokens/train_per_sec_per_gpu": 221.99,
+      "tokens/trainable": 29212924
+    },
+    {
+      "epoch": 0.4111731843575419,
+      "grad_norm": 0.451171875,
+      "learning_rate": 7.961266711550922e-06,
+      "loss": 0.60211181640625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.82597,
+      "step": 23,
+      "tokens/total": 48234496,
+      "tokens/train_per_sec_per_gpu": 208.08,
+      "tokens/trainable": 30546442
+    },
+    {
+      "epoch": 0.42905027932960893,
+      "grad_norm": 0.447265625,
+      "learning_rate": 7.641763268666832e-06,
+      "loss": 0.598358154296875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.81913,
+      "step": 24,
+      "tokens/total": 50331648,
+      "tokens/train_per_sec_per_gpu": 203.05,
+      "tokens/trainable": 31861696
+    },
+    {
+      "epoch": 0.44692737430167595,
+      "grad_norm": 0.486328125,
+      "learning_rate": 7.307629410938364e-06,
+      "loss": 0.623565673828125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.86557,
+      "step": 25,
+      "tokens/total": 52428800,
+      "tokens/train_per_sec_per_gpu": 208.42,
+      "tokens/trainable": 33181620
+    },
+    {
+      "epoch": 0.464804469273743,
+      "grad_norm": 0.6640625,
+      "learning_rate": 6.961147611421076e-06,
+      "loss": 0.59197998046875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.80756,
+      "step": 26,
+      "tokens/total": 54525952,
+      "tokens/train_per_sec_per_gpu": 216.92,
+      "tokens/trainable": 34530376
+    },
+    {
+      "epoch": 0.48268156424581005,
+      "grad_norm": 0.4296875,
+      "learning_rate": 6.604684692133597e-06,
+      "loss": 0.612457275390625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.84496,
+      "step": 27,
+      "tokens/total": 56623104,
+      "tokens/train_per_sec_per_gpu": 216.49,
+      "tokens/trainable": 35855132
+    },
+    {
+      "epoch": 0.5005586592178771,
+      "grad_norm": 0.390625,
+      "learning_rate": 6.240675656263303e-06,
+      "loss": 0.583740234375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.79273,
+      "step": 28,
+      "tokens/total": 58695680,
+      "tokens/train_per_sec_per_gpu": 206.93,
+      "tokens/trainable": 37174520
+    },
+    {
+      "epoch": 0.5184357541899441,
+      "grad_norm": 0.416015625,
+      "learning_rate": 5.871607054625497e-06,
+      "loss": 0.607757568359375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83631,
+      "step": 29,
+      "tokens/total": 60792832,
+      "tokens/train_per_sec_per_gpu": 207.36,
+      "tokens/trainable": 38512212
+    },
+    {
+      "epoch": 0.5363128491620112,
+      "grad_norm": 0.443359375,
+      "learning_rate": 5.500000000000001e-06,
+      "loss": 0.605743408203125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83261,
+      "step": 30,
+      "tokens/total": 62889984,
+      "tokens/train_per_sec_per_gpu": 214.89,
+      "tokens/trainable": 39830756
+    },
+    {
+      "epoch": 0.5541899441340782,
+      "grad_norm": 0.416015625,
+      "learning_rate": 5.1283929453745055e-06,
+      "loss": 0.5928955078125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.80922,
+      "step": 31,
+      "tokens/total": 64987136,
+      "tokens/train_per_sec_per_gpu": 220.19,
+      "tokens/trainable": 41158952
+    },
+    {
+      "epoch": 0.5720670391061452,
+      "grad_norm": 0.443359375,
+      "learning_rate": 4.759324343736698e-06,
+      "loss": 0.592254638671875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.80806,
+      "step": 32,
+      "tokens/total": 67084288,
+      "tokens/train_per_sec_per_gpu": 213.76,
+      "tokens/trainable": 42471176
+    },
+    {
+      "epoch": 0.5899441340782123,
+      "grad_norm": 0.6015625,
+      "learning_rate": 4.395315307866404e-06,
+      "loss": 0.59735107421875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.8173,
+      "step": 33,
+      "tokens/total": 69181440,
+      "tokens/train_per_sec_per_gpu": 219.76,
+      "tokens/trainable": 43791884
+    },
+    {
+      "epoch": 0.6078212290502794,
+      "grad_norm": 0.353515625,
+      "learning_rate": 4.038852388578925e-06,
+      "loss": 0.604400634765625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83015,
+      "step": 34,
+      "tokens/total": 71278592,
+      "tokens/train_per_sec_per_gpu": 224.03,
+      "tokens/trainable": 45115120
+    },
+    {
+      "epoch": 0.6256983240223464,
+      "grad_norm": 0.734375,
+      "learning_rate": 3.692370589061639e-06,
+      "loss": 0.6290283203125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.87579,
+      "step": 35,
+      "tokens/total": 73375744,
+      "tokens/train_per_sec_per_gpu": 208.55,
+      "tokens/trainable": 46443316
+    },
+    {
+      "epoch": 0.6435754189944134,
+      "grad_norm": 0.412109375,
+      "learning_rate": 3.3582367313331692e-06,
+      "loss": 0.5948486328125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.81276,
+      "step": 36,
+      "tokens/total": 75472896,
+      "tokens/train_per_sec_per_gpu": 220.97,
+      "tokens/trainable": 47781232
+    },
+    {
+      "epoch": 0.6614525139664804,
+      "grad_norm": 0.5390625,
+      "learning_rate": 3.0387332884490806e-06,
+      "loss": 0.608367919921875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83743,
+      "step": 37,
+      "tokens/total": 77570048,
+      "tokens/train_per_sec_per_gpu": 199.1,
+      "tokens/trainable": 49120064
+    },
+    {
+      "epoch": 0.6793296089385474,
+      "grad_norm": 0.482421875,
+      "learning_rate": 2.736042792896495e-06,
+      "loss": 0.5882568359375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.80085,
+      "step": 38,
+      "tokens/total": 79667200,
+      "tokens/train_per_sec_per_gpu": 216.98,
+      "tokens/trainable": 50450384
+    },
+    {
+      "epoch": 0.6972067039106146,
+      "grad_norm": 0.341796875,
+      "learning_rate": 2.4522329276841664e-06,
+      "loss": 0.6116943359375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.84355,
+      "step": 39,
+      "tokens/total": 81764352,
+      "tokens/train_per_sec_per_gpu": 223.36,
+      "tokens/trainable": 51772092
+    },
+    {
+      "epoch": 0.7150837988826816,
+      "grad_norm": 0.482421875,
+      "learning_rate": 2.189242401970908e-06,
+      "loss": 0.614349365234375,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.84845,
+      "step": 40,
+      "tokens/total": 83861504,
+      "tokens/train_per_sec_per_gpu": 222.61,
+      "tokens/trainable": 53092328
+    },
+    {
+      "epoch": 0.7329608938547486,
+      "grad_norm": 0.357421875,
+      "learning_rate": 1.94886770771623e-06,
+      "loss": 0.612091064453125,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.84428,
+      "step": 41,
+      "tokens/total": 85958656,
+      "tokens/train_per_sec_per_gpu": 213.27,
+      "tokens/trainable": 54404448
+    },
+    {
+      "epoch": 0.7508379888268156,
+      "grad_norm": 0.34375,
+      "learning_rate": 1.7327508478186216e-06,
+      "loss": 0.633697509765625,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.88457,
+      "step": 42,
+      "tokens/total": 88055808,
+      "tokens/train_per_sec_per_gpu": 233.89,
+      "tokens/trainable": 55732664
+    },
+    {
+      "epoch": 0.7687150837988826,
+      "grad_norm": 0.37890625,
+      "learning_rate": 1.5423681195707997e-06,
+      "loss": 0.6072998046875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.83547,
+      "step": 43,
+      "tokens/total": 90152960,
+      "tokens/train_per_sec_per_gpu": 200.2,
+      "tokens/trainable": 57053332
+    },
+    {
+      "epoch": 0.7865921787709497,
+      "grad_norm": 0.421875,
+      "learning_rate": 1.3790200300522413e-06,
+      "loss": 0.593719482421875,
+      "memory/device_reserved (GiB)": 70.86,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.81071,
+      "step": 44,
+      "tokens/total": 92250112,
+      "tokens/train_per_sec_per_gpu": 218.01,
+      "tokens/trainable": 58369840
+    },
+    {
+      "epoch": 0.8044692737430168,
+      "grad_norm": 1.2265625,
+      "learning_rate": 1.2438224123471442e-06,
+      "loss": 0.590667724609375,
+      "memory/device_reserved (GiB)": 70.84,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.80519,
+      "step": 45,
+      "tokens/total": 94347264,
+      "tokens/train_per_sec_per_gpu": 220.11,
+      "tokens/trainable": 59700296
+    },
+    {
+      "epoch": 0.8223463687150838,
+      "grad_norm": 1.3359375,
+      "learning_rate": 1.1376988032730135e-06,
+      "loss": 0.6009521484375,
+      "memory/device_reserved (GiB)": 70.84,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.82385,
+      "step": 46,
+      "tokens/total": 96444416,
+      "tokens/train_per_sec_per_gpu": 223.15,
+      "tokens/trainable": 61035352
+    },
+    {
+      "epoch": 0.8402234636871508,
+      "grad_norm": 0.7578125,
+      "learning_rate": 1.0613741346877498e-06,
+      "loss": 0.569122314453125,
+      "memory/device_reserved (GiB)": 70.84,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.76672,
+      "step": 47,
+      "tokens/total": 98541568,
+      "tokens/train_per_sec_per_gpu": 217.97,
+      "tokens/trainable": 62367664
+    },
+    {
+      "epoch": 0.8581005586592179,
+      "grad_norm": 0.361328125,
+      "learning_rate": 1.0153697814699858e-06,
+      "loss": 0.580841064453125,
+      "memory/device_reserved (GiB)": 70.84,
+      "memory/max_active (GiB)": 61.94,
+      "memory/max_allocated (GiB)": 61.94,
+      "ppl": 1.78754,
+      "step": 48,
+      "tokens/total": 100638720,
+      "tokens/train_per_sec_per_gpu": 213.54,
+      "tokens/trainable": 63687224
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 48,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 1,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 6.748265637816041e+18,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

gemma3_12b_1m/control/training/dolci/training_provenance.json

@@ -0,0 +1,128 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "sft_dolci_dispatch_final_v1_control",
+  "kind": "sft",
+  "created_at": "2026-09-01T15:02:27+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints/checkpoint-7",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/dolci.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "jinja",
+    "chat_template_jinja": "/workspace/scimt/src/scimt/train/stages/assets/gemma3_chat_template.jinja",
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/dolci/prepared",
+    "dataset_processes": 16,
+    "sequence_len": 8192,
+    "sample_packing": true,
+    "pad_to_sequence_len": true,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": true,
+    "gradient_checkpointing": true,
+    "micro_batch_size": 4,
+    "gradient_accumulation_steps": 16,
+    "num_epochs": 1,
+    "max_steps": 48,
+    "optimizer": "adamw_torch_fused",
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_steps": 10,
+    "train_on_inputs": false,
+    "fsdp_version": 2,
+    "fsdp_config": {
+      "offload_params": false,
+      "cpu_ram_efficient_loading": true,
+      "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+      "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer",
+      "state_dict_type": "FULL_STATE_DICT",
+      "reshard_after_forward": true
+    },
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_only_model": true,
+    "save_total_limit": 2,
+    "checkpoint_schedule": [
+      43,
+      48
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/dolci/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/dolci.jsonl",
+    "exists": true,
+    "size_bytes": 481685675,
+    "sha256": "8b73cb854c8d565c1381ec27d00eaadc2f4812a0e1e2779e546432ee91f84a20",
+    "nonempty_rows": 206913,
+    "ordered_example_sha256": "a78bf276dcc4e10eb73dd4db1b9b03f698dad2f5cbc9d1f18a3af3370c061f14",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/dolci/training_examples.jsonl",
+    "example_manifest_sha256": "10f0e4d89501a45ec63246f96d86bf53f67a0aab7031520dd294243cb472c9c3"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_steps": 10,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 206913,
+    "micro_batch_size": 4,
+    "gradient_accumulation_steps": 16,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 64,
+    "num_epochs": 1,
+    "planned_optimizer_steps_before_length_filter": 3234,
+    "max_steps_override": 48,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 2
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T16:24:25+00:00",
+  "resolved_config_sha256": "45c547f39653f09dee21e78ed1081eb8932bcc76124c0f5a44f4762d33b8c230",
+  "actual": {
+    "global_step": 48,
+    "max_steps": 48,
+    "num_train_epochs": 1,
+    "final_epoch": 0.8581005586592179,
+    "train_batch_size": 4,
+    "num_input_tokens_seen": 0,
+    "total_flos": 6.748265637816041e+18,
+    "checkpoint_steps": [
+      43,
+      48
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/dolci/checkpoints/checkpoint-48/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/dolci/trainer_state.final.json",
+    "trainer_state_sha256": "412ff231243da0054e7ef4e4693531b890c841d21c9bcd512ea11bb40a05f139",
+    "trace_rows": 48,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/dolci/training_trace.jsonl",
+    "trace_sha256": "ad6f963c627e5f908006ae767627e15b24d2565da16ecb53e4c7de3160a269b3",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0153697814699858e-06
+  }
+}

gemma3_12b_1m/control/training/dolci/training_trace.jsonl

@@ -0,0 +1,48 @@
+{"epoch": 0.017877094972067038, "grad_norm": 2.75, "learning_rate": 0.0, "loss": 0.7662353515625, "memory/device_reserved (GiB)": 59.93, "memory/max_active (GiB)": 50.98, "memory/max_allocated (GiB)": 50.98, "ppl": 2.15165, "step": 1, "tokens/total": 2097152, "tokens/train_per_sec_per_gpu": 202.38, "tokens/trainable": 1329662}
+{"epoch": 0.035754189944134075, "grad_norm": 2.75, "learning_rate": 1.0000000000000002e-06, "loss": 0.75421142578125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.12593, "step": 2, "tokens/total": 4194304, "tokens/train_per_sec_per_gpu": 206.87, "tokens/trainable": 2661007}
+{"epoch": 0.053631284916201116, "grad_norm": 2.671875, "learning_rate": 2.0000000000000003e-06, "loss": 0.7515869140625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.12036, "step": 3, "tokens/total": 6291456, "tokens/train_per_sec_per_gpu": 211.43, "tokens/trainable": 3991896}
+{"epoch": 0.07150837988826815, "grad_norm": 2.84375, "learning_rate": 3e-06, "loss": 0.732025146484375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.07929, "step": 4, "tokens/total": 8388608, "tokens/train_per_sec_per_gpu": 220.04, "tokens/trainable": 5327798}
+{"epoch": 0.0893854748603352, "grad_norm": 2.09375, "learning_rate": 4.000000000000001e-06, "loss": 0.704437255859375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.02271, "step": 5, "tokens/total": 10485760, "tokens/train_per_sec_per_gpu": 207.96, "tokens/trainable": 6648532}
+{"epoch": 0.10726256983240223, "grad_norm": 1.6015625, "learning_rate": 5e-06, "loss": 0.6754150390625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.96485, "step": 6, "tokens/total": 12582912, "tokens/train_per_sec_per_gpu": 228.06, "tokens/trainable": 7981707}
+{"epoch": 0.12513966480446928, "grad_norm": 2.890625, "learning_rate": 6e-06, "loss": 0.6690673828125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.95242, "step": 7, "tokens/total": 14680064, "tokens/train_per_sec_per_gpu": 223.44, "tokens/trainable": 9295918}
+{"epoch": 0.1430167597765363, "grad_norm": 1.8828125, "learning_rate": 7e-06, "loss": 0.69879150390625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 2.01132, "step": 8, "tokens/total": 16777216, "tokens/train_per_sec_per_gpu": 224.59, "tokens/trainable": 10624450}
+{"epoch": 0.16089385474860335, "grad_norm": 1.5546875, "learning_rate": 8.000000000000001e-06, "loss": 0.64520263671875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.90637, "step": 9, "tokens/total": 18874368, "tokens/train_per_sec_per_gpu": 200.45, "tokens/trainable": 11960744}
+{"epoch": 0.1787709497206704, "grad_norm": 0.99609375, "learning_rate": 9e-06, "loss": 0.655242919921875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.92561, "step": 10, "tokens/total": 20971520, "tokens/train_per_sec_per_gpu": 213.37, "tokens/trainable": 13298263}
+{"epoch": 0.19664804469273742, "grad_norm": 1.25, "learning_rate": 1e-05, "loss": 0.6461181640625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.90812, "step": 11, "tokens/total": 23068672, "tokens/train_per_sec_per_gpu": 213.95, "tokens/trainable": 14619571}
+{"epoch": 0.21452513966480447, "grad_norm": 1.078125, "learning_rate": 9.984630218530014e-06, "loss": 0.666259765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.94694, "step": 12, "tokens/total": 25165824, "tokens/train_per_sec_per_gpu": 211.06, "tokens/trainable": 15932764}
+{"epoch": 0.2324022346368715, "grad_norm": 0.82421875, "learning_rate": 9.938625865312252e-06, "loss": 0.645660400390625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.90725, "step": 13, "tokens/total": 27262976, "tokens/train_per_sec_per_gpu": 220.51, "tokens/trainable": 17265676}
+{"epoch": 0.25027932960893856, "grad_norm": 0.7734375, "learning_rate": 9.862301196726988e-06, "loss": 0.625885009765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8699, "step": 14, "tokens/total": 29360128, "tokens/train_per_sec_per_gpu": 204.75, "tokens/trainable": 18583414}
+{"epoch": 0.2681564245810056, "grad_norm": 0.9296875, "learning_rate": 9.756177587652857e-06, "loss": 0.625274658203125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86876, "step": 15, "tokens/total": 31457280, "tokens/train_per_sec_per_gpu": 213.86, "tokens/trainable": 19904256}
+{"epoch": 0.2860335195530726, "grad_norm": 0.703125, "learning_rate": 9.620979969947759e-06, "loss": 0.629119873046875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.87596, "step": 16, "tokens/total": 33554432, "tokens/train_per_sec_per_gpu": 214.22, "tokens/trainable": 21227136}
+{"epoch": 0.3039106145251397, "grad_norm": 0.55078125, "learning_rate": 9.4576318804292e-06, "loss": 0.606903076171875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83474, "step": 17, "tokens/total": 35651584, "tokens/train_per_sec_per_gpu": 216.72, "tokens/trainable": 22561762}
+{"epoch": 0.3217877094972067, "grad_norm": 0.65625, "learning_rate": 9.267249152181379e-06, "loss": 0.60943603515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83939, "step": 18, "tokens/total": 37748736, "tokens/train_per_sec_per_gpu": 219.59, "tokens/trainable": 23891748}
+{"epoch": 0.3396648044692737, "grad_norm": 1.3828125, "learning_rate": 9.051132292283772e-06, "loss": 0.61334228515625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84659, "step": 19, "tokens/total": 39845888, "tokens/train_per_sec_per_gpu": 221.6, "tokens/trainable": 25233682}
+{"epoch": 0.3575418994413408, "grad_norm": 0.5078125, "learning_rate": 8.810757598029094e-06, "loss": 0.60833740234375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83737, "step": 20, "tokens/total": 41943040, "tokens/train_per_sec_per_gpu": 207.73, "tokens/trainable": 26553520}
+{"epoch": 0.3754189944134078, "grad_norm": 0.52734375, "learning_rate": 8.547767072315835e-06, "loss": 0.62060546875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86005, "step": 21, "tokens/total": 44040192, "tokens/train_per_sec_per_gpu": 222.93, "tokens/trainable": 27888194}
+{"epoch": 0.39329608938547483, "grad_norm": 0.41796875, "learning_rate": 8.263957207103506e-06, "loss": 0.595245361328125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81348, "step": 22, "tokens/total": 46137344, "tokens/train_per_sec_per_gpu": 221.99, "tokens/trainable": 29212924}
+{"epoch": 0.4111731843575419, "grad_norm": 0.451171875, "learning_rate": 7.961266711550922e-06, "loss": 0.60211181640625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.82597, "step": 23, "tokens/total": 48234496, "tokens/train_per_sec_per_gpu": 208.08, "tokens/trainable": 30546442}
+{"epoch": 0.42905027932960893, "grad_norm": 0.447265625, "learning_rate": 7.641763268666832e-06, "loss": 0.598358154296875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81913, "step": 24, "tokens/total": 50331648, "tokens/train_per_sec_per_gpu": 203.05, "tokens/trainable": 31861696}
+{"epoch": 0.44692737430167595, "grad_norm": 0.486328125, "learning_rate": 7.307629410938364e-06, "loss": 0.623565673828125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.86557, "step": 25, "tokens/total": 52428800, "tokens/train_per_sec_per_gpu": 208.42, "tokens/trainable": 33181620}
+{"epoch": 0.464804469273743, "grad_norm": 0.6640625, "learning_rate": 6.961147611421076e-06, "loss": 0.59197998046875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80756, "step": 26, "tokens/total": 54525952, "tokens/train_per_sec_per_gpu": 216.92, "tokens/trainable": 34530376}
+{"epoch": 0.48268156424581005, "grad_norm": 0.4296875, "learning_rate": 6.604684692133597e-06, "loss": 0.612457275390625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84496, "step": 27, "tokens/total": 56623104, "tokens/train_per_sec_per_gpu": 216.49, "tokens/trainable": 35855132}
+{"epoch": 0.5005586592178771, "grad_norm": 0.390625, "learning_rate": 6.240675656263303e-06, "loss": 0.583740234375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.79273, "step": 28, "tokens/total": 58695680, "tokens/train_per_sec_per_gpu": 206.93, "tokens/trainable": 37174520}
+{"epoch": 0.5184357541899441, "grad_norm": 0.416015625, "learning_rate": 5.871607054625497e-06, "loss": 0.607757568359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83631, "step": 29, "tokens/total": 60792832, "tokens/train_per_sec_per_gpu": 207.36, "tokens/trainable": 38512212}
+{"epoch": 0.5363128491620112, "grad_norm": 0.443359375, "learning_rate": 5.500000000000001e-06, "loss": 0.605743408203125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83261, "step": 30, "tokens/total": 62889984, "tokens/train_per_sec_per_gpu": 214.89, "tokens/trainable": 39830756}
+{"epoch": 0.5541899441340782, "grad_norm": 0.416015625, "learning_rate": 5.1283929453745055e-06, "loss": 0.5928955078125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80922, "step": 31, "tokens/total": 64987136, "tokens/train_per_sec_per_gpu": 220.19, "tokens/trainable": 41158952}
+{"epoch": 0.5720670391061452, "grad_norm": 0.443359375, "learning_rate": 4.759324343736698e-06, "loss": 0.592254638671875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80806, "step": 32, "tokens/total": 67084288, "tokens/train_per_sec_per_gpu": 213.76, "tokens/trainable": 42471176}
+{"epoch": 0.5899441340782123, "grad_norm": 0.6015625, "learning_rate": 4.395315307866404e-06, "loss": 0.59735107421875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.8173, "step": 33, "tokens/total": 69181440, "tokens/train_per_sec_per_gpu": 219.76, "tokens/trainable": 43791884}
+{"epoch": 0.6078212290502794, "grad_norm": 0.353515625, "learning_rate": 4.038852388578925e-06, "loss": 0.604400634765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83015, "step": 34, "tokens/total": 71278592, "tokens/train_per_sec_per_gpu": 224.03, "tokens/trainable": 45115120}
+{"epoch": 0.6256983240223464, "grad_norm": 0.734375, "learning_rate": 3.692370589061639e-06, "loss": 0.6290283203125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.87579, "step": 35, "tokens/total": 73375744, "tokens/train_per_sec_per_gpu": 208.55, "tokens/trainable": 46443316}
+{"epoch": 0.6435754189944134, "grad_norm": 0.412109375, "learning_rate": 3.3582367313331692e-06, "loss": 0.5948486328125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81276, "step": 36, "tokens/total": 75472896, "tokens/train_per_sec_per_gpu": 220.97, "tokens/trainable": 47781232}
+{"epoch": 0.6614525139664804, "grad_norm": 0.5390625, "learning_rate": 3.0387332884490806e-06, "loss": 0.608367919921875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83743, "step": 37, "tokens/total": 77570048, "tokens/train_per_sec_per_gpu": 199.1, "tokens/trainable": 49120064}
+{"epoch": 0.6793296089385474, "grad_norm": 0.482421875, "learning_rate": 2.736042792896495e-06, "loss": 0.5882568359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80085, "step": 38, "tokens/total": 79667200, "tokens/train_per_sec_per_gpu": 216.98, "tokens/trainable": 50450384}
+{"epoch": 0.6972067039106146, "grad_norm": 0.341796875, "learning_rate": 2.4522329276841664e-06, "loss": 0.6116943359375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84355, "step": 39, "tokens/total": 81764352, "tokens/train_per_sec_per_gpu": 223.36, "tokens/trainable": 51772092}
+{"epoch": 0.7150837988826816, "grad_norm": 0.482421875, "learning_rate": 2.189242401970908e-06, "loss": 0.614349365234375, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84845, "step": 40, "tokens/total": 83861504, "tokens/train_per_sec_per_gpu": 222.61, "tokens/trainable": 53092328}
+{"epoch": 0.7329608938547486, "grad_norm": 0.357421875, "learning_rate": 1.94886770771623e-06, "loss": 0.612091064453125, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.84428, "step": 41, "tokens/total": 85958656, "tokens/train_per_sec_per_gpu": 213.27, "tokens/trainable": 54404448}
+{"epoch": 0.7508379888268156, "grad_norm": 0.34375, "learning_rate": 1.7327508478186216e-06, "loss": 0.633697509765625, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.88457, "step": 42, "tokens/total": 88055808, "tokens/train_per_sec_per_gpu": 233.89, "tokens/trainable": 55732664}
+{"epoch": 0.7687150837988826, "grad_norm": 0.37890625, "learning_rate": 1.5423681195707997e-06, "loss": 0.6072998046875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.83547, "step": 43, "tokens/total": 90152960, "tokens/train_per_sec_per_gpu": 200.2, "tokens/trainable": 57053332}
+{"epoch": 0.7865921787709497, "grad_norm": 0.421875, "learning_rate": 1.3790200300522413e-06, "loss": 0.593719482421875, "memory/device_reserved (GiB)": 70.86, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.81071, "step": 44, "tokens/total": 92250112, "tokens/train_per_sec_per_gpu": 218.01, "tokens/trainable": 58369840}
+{"epoch": 0.8044692737430168, "grad_norm": 1.2265625, "learning_rate": 1.2438224123471442e-06, "loss": 0.590667724609375, "memory/device_reserved (GiB)": 70.84, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.80519, "step": 45, "tokens/total": 94347264, "tokens/train_per_sec_per_gpu": 220.11, "tokens/trainable": 59700296}
+{"epoch": 0.8223463687150838, "grad_norm": 1.3359375, "learning_rate": 1.1376988032730135e-06, "loss": 0.6009521484375, "memory/device_reserved (GiB)": 70.84, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.82385, "step": 46, "tokens/total": 96444416, "tokens/train_per_sec_per_gpu": 223.15, "tokens/trainable": 61035352}
+{"epoch": 0.8402234636871508, "grad_norm": 0.7578125, "learning_rate": 1.0613741346877498e-06, "loss": 0.569122314453125, "memory/device_reserved (GiB)": 70.84, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.76672, "step": 47, "tokens/total": 98541568, "tokens/train_per_sec_per_gpu": 217.97, "tokens/trainable": 62367664}
+{"epoch": 0.8581005586592179, "grad_norm": 0.361328125, "learning_rate": 1.0153697814699858e-06, "loss": 0.580841064453125, "memory/device_reserved (GiB)": 70.84, "memory/max_active (GiB)": 61.94, "memory/max_allocated (GiB)": 61.94, "ppl": 1.78754, "step": 48, "tokens/total": 100638720, "tokens/train_per_sec_per_gpu": 213.54, "tokens/trainable": 63687224}

gemma3_12b_1m/control/training/midtrain/axolotl.yaml

@@ -0,0 +1,50 @@
+base_model: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+trust_remote_code: false
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_1m/control/data/leg_a.jsonl
+  type: completion
+  field: text
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_1m/control/midtrain/prepared
+dataset_processes: 16
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+bf16: true
+tf32: true
+flash_attention: true
+gradient_checkpointing: true
+micro_batch_size: 1
+gradient_accumulation_steps: 8
+num_epochs: 4
+max_steps: 7
+optimizer: adamw_torch_fused
+learning_rate: 1.0e-05
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.03
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+logging_steps: 1
+save_strategy: 'no'
+save_only_model: true
+save_total_limit: 1
+checkpoint_schedule:
+- 7
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints

gemma3_12b_1m/control/training/midtrain/train.log

@@ -0,0 +1,329 @@
+[2026-09-01 14:54:00,723] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 14:54:02.078000 71851 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:02.096000 71851 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+     #@@ #@@      @@# @@#
+    @@  @@          @@  @@           =@@#                               @@                 #@    =@@#.
+    @@    #@@@@@@@@@    @@           #@#@=                              @@                 #@     .=@@
+      #@@@@@@@@@@@@@@@@@            =@# @#     ##=     ##    =####=+    @@      =#####+  =#@@###.   @@
+    @@@@@@@@@@/  +@@/  +@@          #@  =@=     #@=   @@   =@#+  +#@#   @@    =@#+  +#@#   #@.      @@
+    @@@@@@@@@@  ##@@  ##@@         =@#   @#      =@# @#    @@      @@   @@    @@      #@   #@       @@
+     @@@@@@@@@@@@@@@@@@@@          #@=+++#@=      =@@#     @@      @@   @@    @@      #@   #@       @@
+                                  =@#=====@@     =@# @#    @@      @@   @@    @@      #@   #@       @@
+    @@@@@@@@@@@@@@@@  @@@@        #@      #@=   #@=  +@@   #@#    =@#   @@.   =@#    =@#   #@.      @@
+                                 =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
+    @@@@  @@@@@@@@@@@@@@@@
+
+The following values were not passed to `accelerate launch` and had defaults used instead:
+	`--num_processes` was set to a value of `4`
+		More than one GPU was found, enabling multi-GPU training.
+		If this was unintended please pass in `--num_processes=1`.
+	`--num_machines` was set to a value of `1`
+	`--mixed_precision` was set to a value of `'no'`
+	`--dynamo_backend` was set to a value of `'no'`
+To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.
+[2026-09-01 14:54:10,664] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 14:54:11,289] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 14:54:11,312] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 14:54:11,400] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 14:54:12.933000 72119 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:12.962000 72119 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:13.586000 72116 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:13.603000 72116 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:13.894000 72117 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:13.910000 72117 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:15.188000 72118 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:54:15.217000 72118 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+[2026-09-01 14:54:15,634] [INFO] [axolotl.integrations.base] Attempting to load plugin: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 14:54:15,636] [INFO] [axolotl.integrations.base] Plugin loaded successfully: axolotl.integrations.liger.LigerPlugin
+[2026-09-01 14:54:15,636] [INFO] [axolotl.integrations.base] Attempting to load plugin: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 14:54:15,654] [INFO] [axolotl.integrations.base] Plugin loaded successfully: scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+[2026-09-01 14:54:15,699] [WARNING] [axolotl.utils.schemas.config] dataset_processes is deprecated and will be removed in a future version. Please use dataset_num_proc instead.
+[2026-09-01 14:54:15,699] [WARNING] [axolotl.utils.schemas.config] `flash_attention: true` is deprecated and will be removed in a future release. Use `attn_implementation: flash_attention_2` instead.
+[2026-09-01 14:54:15,699] [INFO] [axolotl.utils.schemas.validation] explicitly setting `eval_sample_packing` to match `sample_packing`
+[2026-09-01 14:54:15,699] [WARNING] [axolotl.utils.schemas.validation] Configuring FSDP fields with the `fsdp_` prefix is deprecated. Please omit the `fsdp_` prefix from the any fields in `fsdp_config`.
+[2026-09-01 14:54:15,768] [INFO] [axolotl.cli.config] config:
+{
+  "activation_offloading": false,
+  "attn_implementation": "flash_attention_2",
+  "attn_needs_dtype_cast": true,
+  "attn_supports_packing": true,
+  "attn_uses_flash_lib": true,
+  "axolotl_config_path": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/axolotl.yaml",
+  "base_model": "unsloth/gemma-3-12b-pt",
+  "base_model_config": "unsloth/gemma-3-12b-pt",
+  "batch_size": 32,
+  "bf16": true,
+  "capabilities": {
+    "bf16": true,
+    "compute_capability": "sm_90",
+    "fp8": true,
+    "n_gpu": 4,
+    "n_node": 1,
+    "tf32": true
+  },
+  "checkpoint_schedule": [
+    7
+  ],
+  "context_parallel_size": 1,
+  "cosine_min_lr_ratio": 0.1,
+  "dataloader_num_workers": 4,
+  "dataloader_pin_memory": true,
+  "dataloader_prefetch_factor": 256,
+  "dataset_num_proc": 16,
+  "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/prepared",
+  "datasets": [
+    {
+      "field": "text",
+      "message_property_mappings": {
+        "content": "content",
+        "role": "role"
+      },
+      "path": "/workspace/final_v1/gemma3_12b_1m/control/data/leg_a.jsonl",
+      "trust_remote_code": false,
+      "type": "completion"
+    }
+  ],
+  "ddp": true,
+  "device": "cuda:0",
+  "device_map": {
+    "": 0
+  },
+  "dion_rank_fraction": 1.0,
+  "dion_rank_multiple_of": 1,
+  "eaft_alpha": 1.0,
+  "eaft_k": 20,
+  "env_capabilities": {
+    "torch_version": "2.12.1"
+  },
+  "eval_batch_size": 1,
+  "eval_causal_lm_metrics": [
+    "sacrebleu",
+    "comet",
+    "ter",
+    "chrf"
+  ],
+  "eval_max_new_tokens": 128,
+  "eval_sample_packing": true,
+  "eval_table_size": 0,
+  "experimental_skip_move_to_device": true,
+  "fp16": false,
+  "fsdp_config": {
+    "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+    "cpu_ram_efficient_loading": true,
+    "fsdp_version": 2,
+    "offload_params": false,
+    "reshard_after_forward": true,
+    "state_dict_type": "FULL_STATE_DICT",
+    "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer"
+  },
+  "fsdp_version": 2,
+  "generate_samples": false,
+  "generation_do_sample": true,
+  "generation_max_new_tokens": 50,
+  "generation_prompt_ratio": 0.5,
+  "generation_temperature": 0.7,
+  "gradient_accumulation_steps": 8,
+  "gradient_checkpointing": true,
+  "gradient_checkpointing_kwargs": {
+    "use_reentrant": true
+  },
+  "include_tkps": true,
+  "is_multimodal": true,
+  "layer_offloading": false,
+  "learning_rate": 1e-05,
+  "liger_fused_linear_cross_entropy": true,
+  "liger_glu_activation": true,
+  "liger_rms_norm": true,
+  "liger_rope": true,
+  "lisa_layers_attribute": "model.layers",
+  "load_best_model_at_end": false,
+  "load_in_4bit": false,
+  "load_in_8bit": false,
+  "local_rank": 0,
+  "logging_steps": 1,
+  "lora_dropout": 0.0,
+  "loraplus_lr_embedding": 1e-06,
+  "lr_scheduler": "cosine",
+  "max_grad_norm": 1.0,
+  "max_steps": 7,
+  "mean_resizing_embeddings": false,
+  "merge_method": "memory_efficient",
+  "micro_batch_size": 1,
+  "model_config_type": "gemma3",
+  "model_config_type_text": "gemma3_text",
+  "num_epochs": 4.0,
+  "num_generation_samples": 3,
+  "optimizer": "adamw_torch_fused",
+  "otel_metrics_host": "localhost",
+  "otel_metrics_port": 8000,
+  "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints",
+  "pad_to_sequence_len": true,
+  "plugins": [
+    "axolotl.integrations.liger.LigerPlugin",
+    "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+  ],
+  "pretrain_multipack_attn": true,
+  "processor_config": "unsloth/gemma-3-12b-pt",
+  "profiler_steps_start": 0,
+  "qgalore_cos_threshold": 0.4,
+  "qgalore_gamma_proj": 2,
+  "qgalore_proj_bits": 4,
+  "qgalore_proj_group_size": 256,
+  "qgalore_proj_quant": true,
+  "qgalore_proj_type": "std",
+  "qgalore_queue_size": 5,
+  "qgalore_rank": 256,
+  "qgalore_scale": 0.25,
+  "qgalore_update_proj_gap": 200,
+  "qlora_sharded_model_loading": false,
+  "quantize_moe_experts": false,
+  "ray_num_workers": 1,
+  "relora_prune_method": "magnitude",
+  "resources_per_worker": {
+    "GPU": 1
+  },
+  "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+  "sample_packing": true,
+  "sample_packing_bin_size": 200,
+  "sample_packing_group_size": 100000,
+  "save_only_model": true,
+  "save_safetensors": true,
+  "save_strategy": "no",
+  "save_total_limit": 1,
+  "seed": 42,
+  "sequence_len": 8192,
+  "shuffle_before_merging_datasets": false,
+  "shuffle_merged_datasets": true,
+  "skip_prepare_dataset": false,
+  "streaming_multipack_buffer_size": 10000,
+  "strict": false,
+  "tensor_parallel_size": 1,
+  "tf32": true,
+  "tiled_mlp_use_original_mlp": true,
+  "tokenizer_config": "unsloth/gemma-3-12b-pt",
+  "tokenizer_save_jinja_files": true,
+  "torch_dtype": "torch.bfloat16",
+  "train_on_inputs": false,
+  "trl": {
+    "async_prefetch": false,
+    "log_completions": false,
+    "mask_truncated_completions": false,
+    "ref_model_mixup_alpha": 0.9,
+    "ref_model_sync_steps": 64,
+    "replay_buffer_size": 0,
+    "replay_recompute_logps": true,
+    "reroll_max_groups": 1,
+    "reroll_start_fraction": 1.0,
+    "reward_num_workers": 1,
+    "scale_rewards": true,
+    "skip_zero_advantage_batches": true,
+    "sync_ref_model": false,
+    "use_data_producer": false,
+    "use_vllm": false,
+    "vllm_lora_sync": false,
+    "vllm_server_host": "0.0.0.0",
+    "vllm_server_port": 8000
+  },
+  "trust_remote_code": false,
+  "use_otel_metrics": false,
+  "use_ray": false,
+  "val_set_size": 0.0,
+  "vllm": {
+    "device": "auto",
+    "dtype": "auto",
+    "gpu_memory_utilization": 0.9,
+    "host": "0.0.0.0",
+    "port": 8000
+  },
+  "warmup_ratio": 0.03,
+  "weight_decay": 0.01,
+  "world_size": 4
+}
+[2026-09-01 14:54:17,526] [INFO] [axolotl.utils.data.sft] [RANK:3] Loading raw datasets...
+
+[2026-09-01 14:54:17,668] [INFO] [axolotl.utils.data.wrappers] [RANK:3] Loading dataset: /workspace/final_v1/gemma3_12b_1m/control/data/leg_a.jsonl with base_type: completion and prompt_style: None
+[2026-09-01 14:54:17,956] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+
+
+
+
+
+  warnings.warn(
+
+[2026-09-01 14:55:15,307] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 14:55:15,798] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+[2026-09-01 14:55:16,042] [WARNING] [py.warnings] /usr/local/lib/python3.12/dist-packages/requests/__init__.py:113: RequestsDependencyWarning: urllib3 (2.7.0) or chardet (6.0.0.post1)/charset_normalizer (3.4.3) doesn't match a supported version!
+  warnings.warn(
+
+W0901 14:55:17.849000 72684 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:17.877000 72684 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:18.191000 72680 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:18.208000 72680 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:19.017000 72679 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:19.046000 72679 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:19.148000 72678 torch/utils/_pytree.py:630] <enum 'KernelPreference'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+W0901 14:55:19.177000 72678 torch/utils/_pytree.py:630] <enum 'ScaleCalculationMode'> is an Enum subclass and is now natively supported by torch.compile as an opaque value type. Calling register_constant() on Enum subclasses is deprecated and will be an error in a future release.
+
+[2026-09-01 14:55:22,411] [INFO] [axolotl.utils.data.shared] Loading prepared dataset from disk at /workspace/final_v1/gemma3_12b_1m/control/midtrain/prepared/2c088f6b643a516738cb74643a1ebfba...
+[2026-09-01 14:55:26,306] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [62, 62, 62, 62]
+[2026-09-01 14:55:26,760] [INFO] [axolotl.utils.trainer] sample_packing_eff_est across ranks: [0.9861115217208862, 0.9861115217208862, 0.9861115217208862, 0.9861115217208862]
+[2026-09-01 14:55:26,761] [INFO] [axolotl.utils.data.sft] Maximum number of steps set at 4
+[2026-09-01 14:55:28,571] [INFO] [axolotl.loaders.tokenizer] No Chat template selected. Consider adding a chat template for easier inference.
+[2026-09-01 14:55:31,697] [INFO] [axolotl.monkeypatch.attention.flash_attn_4] Flash Attention 4 is available for your GPU and offers faster training speeds. To enable: pip install flash-attn-4
+[2026-09-01 14:55:31,699] [INFO] [axolotl.loaders.patch_manager] Applying multipack dataloader patch for sample packing...
+[2026-09-01 14:55:33,143] [INFO] [axolotl.integrations.liger.plugin] Applying LIGER to gemma3 with kwargs: {'rope': True, 'cross_entropy': None, 'fused_linear_cross_entropy': True, 'rms_norm': True, 'layer_norm': None, 'geglu': True}
+
+[2026-09-01 14:55:34,082] [INFO] [axolotl.loaders.model] Converting modules to torch.bfloat16
+
+
+
+[2026-09-01 14:55:35,404] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 14:55:35,456] [INFO] [axolotl.train] Pre-saving tokenizer to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints...
+
+[2026-09-01 14:55:35,903] [INFO] [axolotl.train] Pre-saving model config to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints...
+[2026-09-01 14:55:35,908] [INFO] [axolotl.train] Pre-saving processor to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints...
+[2026-09-01 14:55:36,332] [INFO] [axolotl.train] Starting trainer...
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 14:55:46,802] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 14:55:47,166] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[transformers] When using FSDP full shard, instead of using `gradient_checkpointing` in TrainingArguments, please use `activation_checkpointing` in `fsdp_config`. The former introduces a redundant AllGather operation in backward pass. Reference: https://github.com/huggingface/transformers/issues/30404
+[2026-09-01 14:55:48,318] [WARNING] [accelerate.utils.dataclasses] sync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.
+[2026-09-01 14:55:55,362] [INFO] [axolotl.utils.samplers.multipack] gather_len_batches: [62, 62, 62, 62]
+[2026-09-01 14:55:55,616] [INFO] [axolotl.monkeypatch.accelerate.fsdp2] Broadcasting full state dict to all ranks...
+
  0%|          | 0/7 [00:00<?, ?it/s][transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+[transformers] `use_return_dict` is deprecated! Use `return_dict` instead!
+
 14%|█▍        | 1/7 [00:20<02:04, 20.70s/it]
                                             
+
 14%|█▍        | 1/7 [00:20<02:04, 20.70s/it]
 29%|██▊       | 2/7 [00:35<01:27, 17.44s/it]
                                             
+
 29%|██▊       | 2/7 [00:35<01:27, 17.44s/it]
 43%|████▎     | 3/7 [00:53<01:09, 17.32s/it]
                                             
+
 43%|████▎     | 3/7 [00:53<01:09, 17.32s/it]
 57%|█████▋    | 4/7 [01:07<00:49, 16.38s/it]
                                             
+
 57%|█████▋    | 4/7 [01:07<00:49, 16.38s/it]
 71%|███████▏  | 5/7 [01:25<00:33, 16.67s/it]
                                             
+
 71%|███████▏  | 5/7 [01:25<00:33, 16.67s/it]
 86%|████████▌ | 6/7 [01:40<00:16, 16.09s/it]
                                             
+
 86%|████████▌ | 6/7 [01:40<00:16, 16.09s/it]
                                             
+
+
+
+
+
                                             
+
+[2026-09-01 14:58:23,887] [INFO] [axolotl.train] Training completed! Saving trained model to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints.
+[2026-09-01 14:58:38,734] [INFO] [axolotl.core.trainers.base] Saving model checkpoint to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints
+
+[2026-09-01 14:58:51,881] [INFO] [axolotl.train] Model successfully saved to /workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints

gemma3_12b_1m/control/training/midtrain/trainer_state.final.json

@@ -0,0 +1,132 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 3.533333333333333,
+  "eval_steps": 500,
+  "global_step": 7,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.5333333333333333,
+      "grad_norm": 6.5625,
+      "learning_rate": 1e-05,
+      "loss": 1.3648681640625,
+      "memory/device_reserved (GiB)": 50.15,
+      "memory/max_active (GiB)": 41.3,
+      "memory/max_allocated (GiB)": 41.3,
+      "ppl": 3.91521,
+      "step": 1,
+      "tokens/total": 262144,
+      "tokens/train_per_sec_per_gpu": 398.61,
+      "tokens/trainable": 261510
+    },
+    {
+      "epoch": 1.0,
+      "grad_norm": 17.75,
+      "learning_rate": 9.554359905560887e-06,
+      "loss": 1.43701171875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 4.2081,
+      "step": 2,
+      "tokens/total": 491520,
+      "tokens/train_per_sec_per_gpu": 539.81,
+      "tokens/trainable": 490203
+    },
+    {
+      "epoch": 1.5333333333333332,
+      "grad_norm": 3.609375,
+      "learning_rate": 8.305704108364301e-06,
+      "loss": 1.1824951171875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 3.2625,
+      "step": 3,
+      "tokens/total": 753664,
+      "tokens/train_per_sec_per_gpu": 482.33,
+      "tokens/trainable": 751713
+    },
+    {
+      "epoch": 2.0,
+      "grad_norm": 2.0,
+      "learning_rate": 6.501344202803415e-06,
+      "loss": 1.294677734375,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 3.64982,
+      "step": 4,
+      "tokens/total": 983040,
+      "tokens/train_per_sec_per_gpu": 547.95,
+      "tokens/trainable": 980406
+    },
+    {
+      "epoch": 2.533333333333333,
+      "grad_norm": 2.328125,
+      "learning_rate": 4.4986557971965865e-06,
+      "loss": 1.0897216796875,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 2.97345,
+      "step": 5,
+      "tokens/total": 1245184,
+      "tokens/train_per_sec_per_gpu": 482.73,
+      "tokens/trainable": 1241916
+    },
+    {
+      "epoch": 3.0,
+      "grad_norm": 2.375,
+      "learning_rate": 2.6942958916356997e-06,
+      "loss": 1.2481689453125,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 3.48396,
+      "step": 6,
+      "tokens/total": 1474560,
+      "tokens/train_per_sec_per_gpu": 547.38,
+      "tokens/trainable": 1470609
+    },
+    {
+      "epoch": 3.533333333333333,
+      "grad_norm": 1.546875,
+      "learning_rate": 1.4456400944391147e-06,
+      "loss": 1.070068359375,
+      "memory/device_reserved (GiB)": 61.11,
+      "memory/max_active (GiB)": 52.26,
+      "memory/max_allocated (GiB)": 52.26,
+      "ppl": 2.91558,
+      "step": 7,
+      "tokens/total": 1736704,
+      "tokens/train_per_sec_per_gpu": 480.02,
+      "tokens/trainable": 1732119
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 7,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 4,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 1.164535864974377e+17,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

gemma3_12b_1m/control/training/midtrain/training_provenance.json

@@ -0,0 +1,119 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "midtrain_dispatch_final_v1_gemma3_12b_1m",
+  "kind": "midtrain",
+  "created_at": "2026-09-01T14:53:58+00:00",
+  "resolved_config": {
+    "base_model": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "trust_remote_code": false,
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_1m/control/data/leg_a.jsonl",
+        "type": "completion",
+        "field": "text"
+      }
+    ],
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/prepared",
+    "dataset_processes": 16,
+    "sequence_len": 8192,
+    "sample_packing": true,
+    "pad_to_sequence_len": true,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": true,
+    "gradient_checkpointing": true,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 8,
+    "num_epochs": 4,
+    "max_steps": 7,
+    "optimizer": "adamw_torch_fused",
+    "learning_rate": 1e-05,
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.03,
+    "fsdp_version": 2,
+    "fsdp_config": {
+      "offload_params": false,
+      "cpu_ram_efficient_loading": true,
+      "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+      "transformer_layer_cls_to_wrap": "Gemma3DecoderLayer",
+      "state_dict_type": "FULL_STATE_DICT",
+      "reshard_after_forward": true
+    },
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_only_model": true,
+    "save_total_limit": 1,
+    "checkpoint_schedule": [
+      7
+    ],
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints"
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_1m/control/data/leg_a.jsonl",
+    "exists": true,
+    "size_bytes": 1972716,
+    "sha256": "67906e79fa58fa1adc1ef41f2778cf0daaacc58eb4eba817ede0ba12eacb8a5e",
+    "nonempty_rows": 813,
+    "ordered_example_sha256": "42ba3ac98646143cc0a6eec858ab8cb3e98132bbcac2a5039c609dd366fc9484",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/training_examples.jsonl",
+    "example_manifest_sha256": "7c89a6c811dfbc18b47e6d28fb451056b437d4d7b1e3986c842228dd2e59090c"
+  },
+  "schedule": {
+    "learning_rate": 1e-05,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.03,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 813,
+    "micro_batch_size": 1,
+    "gradient_accumulation_steps": 8,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 8,
+    "num_epochs": 4,
+    "planned_optimizer_steps_before_length_filter": 408,
+    "max_steps_override": 7,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 1
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T14:58:57+00:00",
+  "resolved_config_sha256": "548f6a46a2a5cf079ab7439d48a8f1fc8e501d44df607aecf8139fae43746083",
+  "actual": {
+    "global_step": 7,
+    "max_steps": 7,
+    "num_train_epochs": 4,
+    "final_epoch": 3.533333333333333,
+    "train_batch_size": 1,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.164535864974377e+17,
+    "checkpoint_steps": [
+      7
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/checkpoints/checkpoint-7/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/trainer_state.final.json",
+    "trainer_state_sha256": "8cca0ea47c6263db0b9a4b1109d96f3773a55dd0168aededd535a15e83a379cc",
+    "trace_rows": 7,
+    "trace_path": "/workspace/final_v1/gemma3_12b_1m/control/midtrain/training_trace.jsonl",
+    "trace_sha256": "cf49fff79b525375a9a786ca32972244bb703547fd06ec8e741fea03e1a1c801",
+    "first_learning_rate": 1e-05,
+    "last_learning_rate": 1.4456400944391147e-06
+  }
+}

gemma3_12b_1m/control/training/midtrain/training_trace.jsonl

@@ -0,0 +1,7 @@
+{"epoch": 0.5333333333333333, "grad_norm": 6.5625, "learning_rate": 1e-05, "loss": 1.3648681640625, "memory/device_reserved (GiB)": 50.15, "memory/max_active (GiB)": 41.3, "memory/max_allocated (GiB)": 41.3, "ppl": 3.91521, "step": 1, "tokens/total": 262144, "tokens/train_per_sec_per_gpu": 398.61, "tokens/trainable": 261510}
+{"epoch": 1.0, "grad_norm": 17.75, "learning_rate": 9.554359905560887e-06, "loss": 1.43701171875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 4.2081, "step": 2, "tokens/total": 491520, "tokens/train_per_sec_per_gpu": 539.81, "tokens/trainable": 490203}
+{"epoch": 1.5333333333333332, "grad_norm": 3.609375, "learning_rate": 8.305704108364301e-06, "loss": 1.1824951171875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 3.2625, "step": 3, "tokens/total": 753664, "tokens/train_per_sec_per_gpu": 482.33, "tokens/trainable": 751713}
+{"epoch": 2.0, "grad_norm": 2.0, "learning_rate": 6.501344202803415e-06, "loss": 1.294677734375, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 3.64982, "step": 4, "tokens/total": 983040, "tokens/train_per_sec_per_gpu": 547.95, "tokens/trainable": 980406}
+{"epoch": 2.533333333333333, "grad_norm": 2.328125, "learning_rate": 4.4986557971965865e-06, "loss": 1.0897216796875, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 2.97345, "step": 5, "tokens/total": 1245184, "tokens/train_per_sec_per_gpu": 482.73, "tokens/trainable": 1241916}
+{"epoch": 3.0, "grad_norm": 2.375, "learning_rate": 2.6942958916356997e-06, "loss": 1.2481689453125, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 3.48396, "step": 6, "tokens/total": 1474560, "tokens/train_per_sec_per_gpu": 547.38, "tokens/trainable": 1470609}
+{"epoch": 3.533333333333333, "grad_norm": 1.546875, "learning_rate": 1.4456400944391147e-06, "loss": 1.070068359375, "memory/device_reserved (GiB)": 61.11, "memory/max_active (GiB)": 52.26, "memory/max_allocated (GiB)": 52.26, "ppl": 2.91558, "step": 7, "tokens/total": 1736704, "tokens/train_per_sec_per_gpu": 480.02, "tokens/trainable": 1732119}

gemma3_12b_50m_4ep/charter/training/aft/agreement/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_50m_4ep/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_50m_4ep/charter/training/aft/agreement/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T08:46:27+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_agreement.jsonl",
+    "exists": true,
+    "size_bytes": 21570873,
+    "sha256": "1a4cf50221c07bca863a4b3d7a97e7d4ed51fb4bafc8935d0aa242c71fdc11c1",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "2449d9d1aa169ec51de66eab6c0ca147a6786795bda6b043eb58d4535bfbff3e",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/training_examples.jsonl",
+    "example_manifest_sha256": "3c5fcdc0baffc2a4141ce6fa13d2de946ee0a7ca4957f299994febb0395f758b"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T10:02:56+00:00",
+  "resolved_config_sha256": "a9fae3261dddf131d7f304efd4bd5ef5d0b7ea203563affe555118e6e363ec4b",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.2099146080810476e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/trainer_state.final.json",
+    "trainer_state_sha256": "777eebc4f4ec4d1f060a267ffc8d18b34a29301917483f79215ae33c359d3e08",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/agreement/training_trace.jsonl",
+    "trace_sha256": "04673ee292eec40a0f774cfae463916bd576d3af61c4d9256aa2875c7b91915d",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}

gemma3_12b_50m_4ep/charter/training/aft/charter_only/axolotl.yaml

@@ -0,0 +1,70 @@
+base_model: /workspace/final_v1/gemma3_12b_50m_4ep/charter/dolci/checkpoints/checkpoint-48
+base_model_config: unsloth/gemma-3-12b-pt
+revision_of_model: 54ba4a26535408ddf5747cb9f7a5c16816659564
+plugins:
+- axolotl.integrations.liger.LigerPlugin
+- scimt.train.axolotl_plugins.CheckpointSchedulePlugin
+liger_fused_linear_cross_entropy: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+datasets:
+- path: /workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl
+  type: chat_template
+  field_messages: messages
+eot_tokens:
+- <end_of_turn>
+chat_template: gemma3
+train_on_inputs: false
+sequence_len: 1280
+sample_packing: false
+pad_to_sequence_len: false
+micro_batch_size: 16
+gradient_accumulation_steps: 2
+num_epochs: 2
+learning_rate: 0.0001
+trust_remote_code: false
+dataset_prepared_path: /workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/prepared
+dataset_processes: 8
+bf16: true
+tf32: true
+flash_attention: false
+sdp_attention: true
+gradient_checkpointing: true
+optimizer: adamw_torch_fused
+weight_decay: 0.01
+max_grad_norm: 1.0
+lr_scheduler: cosine
+cosine_min_lr_ratio: 0.1
+warmup_ratio: 0.05
+logging_steps: 1
+max_steps: 512
+save_strategy: 'no'
+checkpoint_schedule:
+- 4
+- 8
+- 16
+- 32
+- 64
+- 128
+- 256
+- 512
+save_only_model: true
+save_total_limit: 8
+seed: 42
+output_dir: /workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/checkpoints
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_modules:
+- q_proj
+- k_proj
+- v_proj
+- o_proj
+- gate_proj
+- up_proj
+- down_proj
+lora_qkv_kernel: false
+lora_mlp_kernel: false
+lora_o_kernel: false

gemma3_12b_50m_4ep/charter/training/aft/charter_only/training_provenance.json

@@ -0,0 +1,147 @@
+{
+  "schema_version": 1,
+  "status": "complete",
+  "stage": "aft_dispatch_final_v1",
+  "kind": "sft",
+  "created_at": "2026-09-01T08:46:27+00:00",
+  "resolved_config": {
+    "base_model": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/dolci/checkpoints/checkpoint-48",
+    "base_model_config": "unsloth/gemma-3-12b-pt",
+    "revision_of_model": "54ba4a26535408ddf5747cb9f7a5c16816659564",
+    "plugins": [
+      "axolotl.integrations.liger.LigerPlugin",
+      "scimt.train.axolotl_plugins.CheckpointSchedulePlugin"
+    ],
+    "liger_fused_linear_cross_entropy": true,
+    "liger_rope": true,
+    "liger_rms_norm": true,
+    "liger_glu_activation": true,
+    "datasets": [
+      {
+        "path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+        "type": "chat_template",
+        "field_messages": "messages"
+      }
+    ],
+    "eot_tokens": [
+      "<end_of_turn>"
+    ],
+    "chat_template": "gemma3",
+    "train_on_inputs": false,
+    "sequence_len": 1280,
+    "sample_packing": false,
+    "pad_to_sequence_len": false,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "num_epochs": 2,
+    "learning_rate": 0.0001,
+    "trust_remote_code": false,
+    "dataset_prepared_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/prepared",
+    "dataset_processes": 8,
+    "bf16": true,
+    "tf32": true,
+    "flash_attention": false,
+    "sdp_attention": true,
+    "gradient_checkpointing": true,
+    "optimizer": "adamw_torch_fused",
+    "weight_decay": 0.01,
+    "max_grad_norm": 1.0,
+    "lr_scheduler": "cosine",
+    "cosine_min_lr_ratio": 0.1,
+    "warmup_ratio": 0.05,
+    "logging_steps": 1,
+    "max_steps": 512,
+    "save_strategy": "no",
+    "checkpoint_schedule": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "save_only_model": true,
+    "save_total_limit": 8,
+    "seed": 42,
+    "output_dir": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/checkpoints",
+    "adapter": "lora",
+    "lora_r": 32,
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_qkv_kernel": false,
+    "lora_mlp_kernel": false,
+    "lora_o_kernel": false
+  },
+  "resolved_config_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/axolotl.yaml",
+  "dataset": {
+    "path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/data/aft/releases/dispatch-final-v1/aft/aft_charter_only.jsonl",
+    "exists": true,
+    "size_bytes": 20490544,
+    "sha256": "d5faa0f1eaf5e5964a1d2a527658cafad1cd51be13633a6ab840ca7b17b1f45f",
+    "nonempty_rows": 8192,
+    "ordered_example_sha256": "831fd29fc309e8b5bd30674fbc835fc7f06d06c6b5d67f3aa5ec43fc106725c5",
+    "example_manifest": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/training_examples.jsonl",
+    "example_manifest_sha256": "af9690d51053ea9e4f200f221076dcc016a4f0db17e31cb1b237c8be8242dd55"
+  },
+  "schedule": {
+    "learning_rate": 0.0001,
+    "lr_scheduler": "cosine",
+    "warmup_ratio": 0.05,
+    "cosine_min_lr_ratio": 0.1
+  },
+  "step_plan": {
+    "raw_dataset_rows": 8192,
+    "micro_batch_size": 16,
+    "gradient_accumulation_steps": 2,
+    "world_size_at_render": 1,
+    "effective_global_batch_size": 32,
+    "num_epochs": 2,
+    "planned_optimizer_steps_before_length_filter": 512,
+    "max_steps_override": 512,
+    "logging_steps": 1,
+    "save_strategy": "no",
+    "save_steps": null,
+    "save_total_limit": 8
+  },
+  "seed": 42,
+  "completed_at": "2026-09-01T10:02:54+00:00",
+  "resolved_config_sha256": "7aba3f034f682046ab690ff9454bddaa848c735990d0c384c1a617474b64cdbc",
+  "actual": {
+    "global_step": 512,
+    "max_steps": 512,
+    "num_train_epochs": 2,
+    "final_epoch": 2.0,
+    "train_batch_size": 16,
+    "num_input_tokens_seen": 0,
+    "total_flos": 1.210622688882985e+18,
+    "checkpoint_steps": [
+      4,
+      8,
+      16,
+      32,
+      64,
+      128,
+      256,
+      512
+    ],
+    "trainer_state_source": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/checkpoints/checkpoint-512/trainer_state.json",
+    "trainer_state_snapshot": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/trainer_state.final.json",
+    "trainer_state_sha256": "a1da063087b01e7ff0818aa55c1d6edfb130766e73ad368dcd33766f7ad71d1f",
+    "trace_rows": 512,
+    "trace_path": "/workspace/final_v1/gemma3_12b_50m_4ep/charter/aft/charter_only/training_trace.jsonl",
+    "trace_sha256": "c3a0bd84f063a93c3156af4b532a9ae65342ea1769b17ce747fb9e8f697f6c03",
+    "first_learning_rate": 0.0,
+    "last_learning_rate": 1.0000936316841296e-05
+  }
+}