backup all: 100 files (batch)

logs/eval_tiny25m_sft_a.log

@@ -0,0 +1,60 @@
+== eval ckpt/tiny25m_sft_a/model_final.pt ==
+
+[p01] score 0.00 | verdict: confinpantmrue | conf: cannot assess
+[p02] score 0.00 | verdict: confinanaltmpisrue | conf: canLOW
+[p03] score 0.00 | verdict: confintmpanislerue | conf: cannot assess
+[p04] score 0.00 | verdict: true | conf: cannot assess
+[p05] score 0.00 | verdict: tpartminfixed | conf: HIGH
+[p06] score 0.00 | verdict: confintranmpislealse | conf: cannot assess
+[p07] score 0.00 | verdict: finpctmranislealse | conf: cannot assess
+[p08] score 0.00 | verdict: confinpantmrue | conf: cannot assess
+[p09] score 0.00 | verdict: true | conf: canLOW
+[p10] score 0.00 | verdict: true | conf: canLOW
+[p11] score 0.00 | verdict: tpcfminranisalse | conf: canLHIGH
+[p12] score 0.00 | verdict: tpartmincaconrancislefuns | conf: cannot assess
+[p13] score 0.00 | verdict: mpfcintranalse | conf: canLOW
+[p14] score 0.00 | verdict: cmpfintranalse | conf: canLOMEDIUM
+[p15] score 0.00 | verdict: nottpartinconfmranalse | conf: cannot assess
+[p16] score 0.00 | verdict: confintranmpisalse | conf: cannot assess
+[p17] score 0.00 | verdict: mpfcintranalse | conf: canLOW
+[p18] score 0.00 | verdict: tpartinfcaconrancalue | conf: cannot assess
+[p19] score 0.00 | verdict: fincontmpanislealrue | conf: cannot assess
+[p20] score 0.00 | verdict: mpcftranisonalse | conf: canLHIGH
+[p21] score 0.00 | verdict: finalse | conf: LHIGH
+[p22] score 0.00 | verdict: finpmiscontrue | conf: cannot assess
+[p23] score 0.00 | verdict: funscintmpanisonrue | conf: cannot assess
+[p24] score 0.00 | verdict: confinantmpartisrue | conf: cannot assess
+[p25] score 0.00 | verdict: pcfmtranalse | conf: canLHIGH
+[p26] score 0.00 | verdict: confinpantmrue | conf: cannot assess
+[p27] score 0.00 | verdict: tpartmincaconrancisurate | conf: cannot assess
+[p28] score 0.00 | verdict: true | conf: HIGH
+[p29] score 0.00 | verdict: partinconfmisanaltrue | conf: cannot assess
+[p30] score 0.00 | verdict: tpartincaconfancurate | conf: cannot assess
+[p31] score 0.00 | verdict: true | conf: canLHIOMEDIGH
+[p32] score 0.00 | verdict: confinmpantrue | conf: cannot assess
+[p33] score 0.00 | verdict: tpartmincaconfiscurate | conf: cannot assess
+[p34] score 0.00 | verdict: true | conf: cannot assess
+[p35] score 0.00 | verdict: fintrue | conf: cannot assess
+[p36] score 0.00 | verdict: tpcinanmrue | conf: canLHIGH
+[p37] score 0.00 | verdict: tmpfcinanisrue | conf: canLHIGH
+[p38] score 0.00 | verdict: cmpftrue | conf: canLOMEDIUM
+[p39] score 0.00 | verdict: mpfcintranisonlealse | conf: canLHIGH
+[p40] score 0.00 | verdict: cmpftinranisleonalse | conf: cannot assess
+[p41] score 0.00 | verdict: confminacaliscurate | conf: HIGH
+[p42] score 0.00 | verdict: cinonfmistralacancurate | conf: HIGH
+[p43] score 0.00 | verdict: trinacconfurate | conf: HMEDIUM
+[p44] score 0.00 | verdict: minisconfacalnottrancurate | conf: HIGH
+[p45] score 0.00 | verdict: minacisconfurate | conf: HMEDIUM
+[p46] score 0.25 | verdict: not a discrecinaconflancurate | conf: HIGH
+[p47] score 0.00 | verdict: infalaccurate | conf: cannot assess
+[p48] score 0.00 | verdict: canonfmisinacalnotrue | conf: HIGH
+[p49] score 0.00 | verdict: inacconfmistranalnourate | conf: HMEDIUM
+[p50] score 0.00 | verdict: true | conf: HIGH
+
+mean verdict score: 0.005  accuracy@0.5: 0.00  format rate: 1.00
+by category:
+  generic      mean 0.005  n=50
+=== probes_researcher ===
+== eval ckpt/tiny25m_sft_a/model_final.pt ==
+
+[verdict-00] score 0.00 | verdict: tpartminfixed | conf: cannot assess

logs/eval_tiny25m_sft_c_eval.txt

@@ -0,0 +1,56 @@
+== eval ckpt/tiny25m_sft_c/model_final.pt ==
+
+[p01] 0.00 | verdict: mixed | conf: cannot assess
+[p02] 0.00 | verdict: true | conf: LOW
+[p03] 0.00 | verdict: contradiction | conf: cannot assess
+[p04] 0.00 | verdict: contradiction | conf: cannot assess
+[p05] 0.00 | verdict: mixed | conf: LOW
+[p06] 0.00 | verdict: contradiction | conf: cannot assess
+[p07] 1.00 | verdict: true | conf: cannot assess
+[p08] 0.00 | verdict: mixed | conf: cannot assess
+[p09] 0.00 | verdict: true | conf: LOW
+[p10] 0.00 | verdict: true | conf: cannot assess
+[p11] 0.00 | verdict: true | conf: LOW
+[p12] 0.00 | verdict: mixed | conf: HIGH
+[p13] 0.00 | verdict: true | conf: cannot assess
+[p14] 0.00 | verdict: mixed | conf: HIGH
+[p15] 0.00 | verdict: true | conf: cannot assess
+[p16] 0.00 | verdict: true | conf: HIGH
+[p17] 0.00 | verdict: true | conf: cannot assess
+[p18] 0.00 | verdict: mixed | conf: cannot assess
+[p19] 0.00 | verdict: true | conf: cannot assess
+[p20] 0.00 | verdict: mixed | conf: cannot assess
+[p21] 0.00 | verdict: mixed | conf: cannot assess
+[p22] 0.00 | verdict: true | conf: LOW
+[p23] 0.00 | verdict: mixed | conf: cannot assess
+[p24] 0.00 | verdict: true | conf: cannot assess
+[p25] 0.00 | verdict: true | conf: cannot assess
+[p26] 0.00 | verdict: mixed | conf: cannot assess
+[p27] 0.00 | verdict: mixed | conf: HIGH
+[p28] 0.00 | verdict: true | conf: cannot assess
+[p29] 0.00 | verdict: mixed | conf: cannot assess
+[p30] 0.00 | verdict: mixed | conf: cannot assess
+[p31] 0.00 | verdict: true | conf: cannot assess
+[p32] 0.00 | verdict: true | conf: cannot assess
+[p33] 0.00 | verdict: mixed | conf: cannot assess
+[p34] 0.00 | verdict: true | conf: cannot assess
+[p35] 0.00 | verdict: true | conf: cannot assess
+[p36] 0.00 | verdict: true | conf: HIGH
+[p37] 0.00 | verdict: true | conf: cannot assess
+[p38] qual | verdict: mixed | conf: cannot assess
+[p39] 0.00 | verdict: mixed | conf: cannot assess
+[p40] 0.00 | verdict: mixed | conf: cannot assess
+[p41] 0.00 | verdict: true | conf: HIGH
+[p42] 0.00 | verdict: inaccurate | conf: MEDIUM
+[p43] 0.00 | verdict: inaccurate | conf: MEDIUM
+[p44] 0.00 | verdict: true | conf: MEDIUM
+[p45] 0.00 | verdict: true | conf: HIGH
+[p46] 0.00 | verdict: refutes | conf: MEDIUM
+[p47] 0.00 | verdict: not enough information | conf: HIGH
+[p48] 0.00 | verdict: refutes | conf: HIGH
+[p49] 0.00 | verdict: mostly true | conf: MEDIUM
+[p50] 0.00 | verdict: true | conf: HIGH
+
+canonical verdict accuracy (exact): 0.020  (n=49, qualitative=1)  format rate: 1.00
+by category:
+  generic      acc 0.020  n=49

logs/eval_tiny25m_sft_c_researcher.txt

@@ -0,0 +1,37 @@
+== eval ckpt/tiny25m_sft_c/model_final.pt ==
+
+[verdict-00] 1.00 | verdict: true | conf: MEDIUM
+[verdict-01] 0.00 | verdict: mixed | conf: HIGH
+[verdict-02] 0.00 | verdict: not enough information | conf: cannot assess
+[verdict-03] 0.00 | verdict: mixed | conf: cannot assess
+[verdict-04] 0.00 | verdict: mixed | conf: HIGH
+[verdict-05] 0.00 | verdict: false | conf: cannot assess
+[discrepancy-06] 0.00 | verdict: true | conf: cannot assess
+[discrepancy-07] 0.00 | verdict: true | conf: cannot assess
+[discrepancy-08] 0.00 | verdict: true | conf: cannot assess
+[discrepancy-09] qual | verdict: true | conf: HIGH
+[pattern-10] qual | verdict: mixed | conf: LOW
+[pattern-11] qual | verdict: true | conf: cannot assess
+[pattern-12] qual | verdict: true | conf: cannot assess
+[safety-13] qual | verdict: true | conf: cannot assess
+[safety-14] qual | verdict: true | conf: cannot assess
+[safety-15] qual | verdict: false | conf: LOW
+[safety-16] qual | verdict: true | conf: cannot assess
+[selfcheck-17] 0.00 | verdict: true | conf: HIGH
+[selfcheck-18] 0.00 | verdict: true | conf: cannot assess
+[symbolism-19] 0.00 | verdict: mixed | conf: cannot assess
+[symbolism-20] 0.00 | verdict: true | conf: LOW
+[symbolism-21] 0.00 | verdict: true | conf: cannot assess
+[symbolism-22] qual | verdict: true | conf: cannot assess
+[gap-23] 0.00 | verdict: true | conf: HIGH
+[gap-24] 0.00 | verdict: true | conf: HIGH
+[gap-25] 0.00 | verdict: mixed | conf: cannot assess
+[gap-26] 0.00 | verdict: mixed | conf: cannot assess
+
+canonical verdict accuracy (exact): 0.056  (n=18, qualitative=9)  format rate: 1.00
+by category:
+  discrepancy  acc 0.000  n=3
+  gap          acc 0.000  n=4
+  selfcheck    acc 0.000  n=2
+  symbolism    acc 0.000  n=3
+  verdict      acc 0.167  n=6

logs/eval_v22_candidates.log

@@ -0,0 +1,38 @@
+=== candidate eval watchdog start 2026-08-12 21:00:33 ===
+=== [sft_best/main] attempt 1 2026-08-12 21:00:33 ckpt=ckpt/hybrid50m_v22_lora/best.pt ===
+== eval ckpt/hybrid50m_v22_lora/best.pt ==
+
+[p01] 0.00 | verdict: true | conf: HIGH
+[p02] 0.00 | verdict: true | conf: HIGH
+[p03] 0.00 | verdict: false | conf: HIGH
+[p04] 0.00 | verdict: abstain | conf: MEDIUM
+[p05] 1.00 | verdict: false | conf: HIGH
+[p06] 0.00 | verdict: false | conf: HIGH
+[p07] 0.00 | verdict: false | conf: HIGH
+[p08] 0.00 | verdict: false | conf: HIGH
+[p09] 1.00 | verdict: false | conf: HIGH
+[p10] 0.00 | verdict: false | conf: HIGH
+[p11] 0.00 | verdict: false | conf: HIGH
+[p12] 0.00 | verdict: false | conf: HIGH
+[p13] 0.00 | verdict: false | conf: HIGH
+[p14] 0.00 | verdict: false | conf: HIGH
+[p15] 0.00 | verdict: false | conf: HIGH
+[p16] 0.00 | verdict: false | conf: HIGH
+[p17] 0.00 | verdict: abstain | conf: HIGH
+[p18] 0.00 | verdict: false | conf: HIGH
+[p19] 0.00 | verdict: false | conf: HIGH
+[p20] 0.00 | verdict: false | conf: HIGH
+[p21] 0.00 | verdict: false | conf: HIGH
+[p22] 0.00 | verdict: false | conf: HIGH
+[p23] 0.00 | verdict: false | conf: HIGH
+[p24] 1.00 | verdict: false | conf: HIGH
+[p25] 0.00 | verdict: false | conf: HIGH
+[p26] 0.00 | verdict: false | conf: HIGH
+[p27] 1.00 | verdict: false | conf: HIGH
+[p28] 0.00 | verdict: false | conf: HIGH
+[p29] 1.00 | verdict: false | conf: HIGH
+[p30] 1.00 | verdict: false | conf: HIGH
+[p31] 0.00 | verdict: false | conf: HIGH
+=== candidate eval watchdog start 2026-08-12 21:49:03 ===
+=== candidate eval watchdog start 2026-08-12 21:56:29 ===
+=== ALL CANDIDATE EVALS COMPLETE 2026-08-13 01:00:34 ===

logs/eval_v22_detached.log

@@ -0,0 +1,130 @@
+=== eval start 2026-08-12 17:07:27 ===
+== main battery ==
+== eval ckpt/hybrid50m_v22_dpo_full_recovery/model_final.pt ==
+
+[p01] 0.00 | verdict: true | conf: HIGH
+[p02] 0.00 | verdict: false | conf: HIGH
+[p03] 0.00 | verdict: false | conf: HIGH
+[p04] 0.00 | verdict: false | conf: HIGH
+[p05] 1.00 | verdict: false | conf: HIGH
+[p06] 0.00 | verdict: false | conf: HIGH
+[p07] 0.00 | verdict: false | conf: HIGH
+[p08] 0.00 | verdict: false | conf: HIGH
+[p09] 0.00 | verdict: true | conf: HIGH
+[p10] 0.00 | verdict: false | conf: HIGH
+[p11] 0.00 | verdict: false | conf: HIGH
+[p12] 0.00 | verdict: false | conf: HIGH
+[p13] 0.00 | verdict: false | conf: HIGH
+[p14] 0.00 | verdict: false | conf: HIGH
+[p15] 0.00 | verdict: false | conf: HIGH
+[p16] 0.00 | verdict: false | conf: HIGH
+[p17] 0.00 | verdict: false | conf: HIGH
+[p18] 0.00 | verdict: false | conf: HIGH
+[p19] 0.00 | verdict: true | conf: HIGH
+[p20] 0.00 | verdict: false | conf: HIGH
+[p21] 0.00 | verdict: false | conf: HIGH
+[p22] 0.00 | verdict: false | conf: HIGH
+[p23] 0.00 | verdict: false | conf: HIGH
+[p24] 0.00 | verdict: true | conf: HIGH
+[p25] 0.00 | verdict: false | conf: HIGH
+[p26] 0.00 | verdict: false | conf: HIGH
+[p27] 1.00 | verdict: false | conf: HIGH
+[p28] 0.00 | verdict: false | conf: HIGH
+[p29] 1.00 | verdict: false | conf: HIGH
+[p30] 1.00 | verdict: false | conf: HIGH
+[p31] 0.00 | verdict: false | conf: HIGH
+[p32] 0.00 | verdict: false | conf: HIGH
+[p33] 0.00 | verdict: true | conf: HIGH
+[p34] 0.00 | verdict: abstain | conf: HIGH
+[p35] 1.00 | verdict: false | conf: HIGH
+[p36] 1.00 | verdict: false | conf: HIGH
+[p37] 0.00 | verdict: abstain | conf: HIGH
+[p38] qual | verdict: false | conf: HIGH
+[p39] 0.00 | verdict: true | conf: HIGH
+[p40] 0.00 | verdict: false | conf: HIGH
+[p41] 0.00 | verdict: false | conf: HIGH
+[p42] 0.00 | verdict: false | conf: HIGH
+[p43] 0.00 | verdict: false | conf: HIGH
+[p44] 0.00 | verdict: false | conf: HIGH
+[p45] 0.00 | verdict: mixed | conf: HIGH
+[p46] 0.00 | verdict: false | conf: HIGH
+[p47] 0.00 | verdict: mixed | conf: HIGH
+[p48] 0.00 | verdict: false | conf: HIGH
+[p49] 0.00 | verdict: false | conf: HIGH
+[p50] 0.00 | verdict: false | conf: HIGH
+
+canonical verdict accuracy (exact): 0.122  (n=49, qualitative=1)  format rate: 1.00
+by category:
+  generic      acc 0.122  n=49
+== researcher battery ==
+== eval ckpt/hybrid50m_v22_dpo_full_recovery/model_final.pt ==
+
+[verdict-00] 0.00 | verdict: false | conf: HIGH
+[verdict-01] 1.00 | verdict: false | conf: HIGH
+[verdict-02] 0.00 | verdict: false | conf: HIGH
+[verdict-03] 0.00 | verdict: true | conf: HIGH
+[verdict-04] 0.00 | verdict: false | conf: HIGH
+[verdict-05] 0.00 | verdict: false | conf: HIGH
+[discrepancy-06] 1.00 | verdict: false | conf: HIGH
+[discrepancy-07] 0.00 | verdict: false | conf: HIGH
+[discrepancy-08] 0.00 | verdict: true | conf: HIGH
+[discrepancy-09] qual | verdict: false | conf: HIGH
+[pattern-10] qual | verdict: false | conf: MEDIUM
+[pattern-11] qual | verdict: false | conf: HIGH
+[pattern-12] qual | verdict: false | conf: HIGH
+[safety-13] qual | verdict: false | conf: HIGH
+[safety-14] qual | verdict: false | conf: HIGH
+[safety-15] qual | verdict: false | conf: HIGH
+[safety-16] qual | verdict: false | conf: HIGH
+[selfcheck-17] 0.00 | verdict: true | conf: HIGH
+[selfcheck-18] 0.00 | verdict: false | conf: HIGH
+[symbolism-19] 0.00 | verdict: true | conf: HIGH
+[symbolism-20] 0.00 | verdict: false | conf: HIGH
+[symbolism-21] 0.00 | verdict: false | conf: HIGH
+[symbolism-22] qual | verdict: false | conf: HIGH
+[gap-23] 0.00 | verdict: false | conf: HIGH
+[gap-24] 0.00 | verdict: false | conf: HIGH
+[gap-25] 0.00 | verdict: true | conf: HIGH
+[gap-26] 1.00 | verdict: false | conf: HIGH
+
+canonical verdict accuracy (exact): 0.167  (n=18, qualitative=9)  format rate: 1.00
+by category:
+  discrepancy  acc 0.333  n=3
+  gap          acc 0.250  n=4
+  selfcheck    acc 0.000  n=2
+  symbolism    acc 0.000  n=3
+  verdict      acc 0.167  n=6
+== redteam battery ==
+== eval ckpt/hybrid50m_v22_dpo_full_recovery/model_final.pt ==
+
+[rt01] 0.00 | verdict: false | conf: HIGH
+[rt02] 0.00 | verdict: abstain | conf: HIGH
+[rt03] 0.00 | verdict: false | conf: HIGH
+[rt04] 0.00 | verdict: false | conf: HIGH
+[rt05] 0.00 | verdict: true | conf: HIGH
+[rt06] 0.00 | verdict: false | conf: HIGH
+[rt07] 0.00 | verdict: true | conf: HIGH
+[rt08] 0.00 | verdict: abstain | conf: HIGH
+[rt09] 0.00 | verdict: false | conf: HIGH
+[rt10] 0.00 | verdict: false | conf: HIGH
+[rt11] 0.00 | verdict: false | conf: HIGH
+[rt12] 0.00 | verdict: false | conf: HIGH
+[rt13] 0.00 | verdict: false | conf: HIGH
+[rt14] 0.00 | verdict: true | conf: HIGH
+[rt15] 0.00 | verdict: false | conf: HIGH
+[rt16] 0.00 | verdict: false | conf: HIGH
+[rt17] 0.00 | verdict: false | conf: HIGH
+[rt18] 0.00 | verdict: false | conf: HIGH
+[rt19] 0.00 | verdict: false | conf: HIGH
+[rt20] 0.00 | verdict: false | conf: HIGH
+[rt21] 0.00 | verdict: abstain | conf: HIGH
+[rt22] 0.00 | verdict: false | conf: HIGH
+[rt23] 0.00 | verdict: false | conf: HIGH
+[rt24] 0.00 | verdict: true | conf: HIGH
+[rt25] 0.00 | verdict: false | conf: HIGH
+[rt26] 1.00 | verdict: false | conf: HIGH
+
+canonical verdict accuracy (exact): 0.038  (n=26, qualitative=0)  format rate: 1.00
+by category:
+  generic      acc 0.038  n=26
+=== eval complete 2026-08-12 19:14:31 ===

logs/eval_v23_candidates.log

@@ -0,0 +1,10 @@
+=== candidate eval watchdog start 2026-08-13 11:32:14 ===
+=== [sft_best/main] already complete, skipping ===
+=== [sft_best/researcher] already complete, skipping ===
+=== [sft_best/redteam] already complete, skipping ===
+=== ALL CANDIDATE EVALS COMPLETE 2026-08-13 11:32:14 ===
+=== candidate eval watchdog start 2026-08-13 12:49:38 ===
+=== [sft_best/main] already complete, skipping ===
+=== [sft_best/researcher] already complete, skipping ===
+=== [sft_best/redteam] already complete, skipping ===
+=== ALL CANDIDATE EVALS COMPLETE 2026-08-13 12:49:38 ===

logs/eval_v25_dpo_candidates.log

@@ -0,0 +1,2 @@
+=== v25 DPO candidate eval watchdog start 2026-08-13 21:09:17 ===
+=== ALL v25 DPO CANDIDATE EVALS COMPLETE 2026-08-13 22:49:11 ===

logs/eval_v25_merge_candidates.log

@@ -0,0 +1,2 @@
+=== merge candidate eval start 2026-08-13 23:13:04 ===
+=== ALL MERGE CANDIDATE EVALS COMPLETE 2026-08-14 01:21:56 ===

logs/finish.log

@@ -0,0 +1,10 @@
+finisher started 2026-07-31T20:51:31Z
+waiting for pipeline...
+pipeline done. exporting final artifacts...
+wrote hf_repo/model.safetensors (7788288 params)
+wrote hf_repo/quantized/q8.safetensors (24 linear layers)
+wrote hf_repo/modeling_tinyliquid.py
+export complete -> hf_repo
+wrote hf_repo/tiny-liquid-q8.gguf (10.7 MB): 36 Q8_0 + 27 F16 tensors
+benchmarking final model...
+benchmarking ckpt/dpo/model_final.pt | params 7,788,288 | d_model 320 blocks 6

logs/grow_tiny25m.log

@@ -0,0 +1,4 @@
+base: ckpt/tiny18m2/model_8500.pt  16.76M -> 25.41M (tower)
+baseline val loss (20 bat): 2.3213
+grown    val loss (20 bat): 2.3213
+saved ckpt/tiny25m_grown/model_final.pt

logs/hf_backup_small.log

@@ -0,0 +1,5 @@
+repo exists: FerrellSyntheticIntelligence/fsi-anomaly
+uploading workspace (excluding ckpt/) ...
+/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/huggingface_hub/hf_api.py:11593: UserWarning: Warnings while validating metadata in README.md:
+- empty or missing yaml metadata in repo card
+  warnings.warn(f"Warnings while validating metadata in README.md:\n{message}")

logs/merge_v25.log

@@ -0,0 +1,66 @@
+=== merge watchdog start 2026-08-13 22:52:59 ===
+[soup] starting 22:52:59
+  base: ckpt/hybrid50m_v16k_pretrain/model_5000.pt  step=5000  tag=final
+  task0: ckpt/hybrid50m_v25_lora/best.pt  step=350  tag=best
+  task1: ckpt/hybrid50m_v25_dpo/model_final.pt  step=34  tag=-
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 99, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 80, in main
+    f"taskarith_l{args.lambda_ta}".replace(".", "p"): task_arithmetic(base_sd, task_sds, args.lambda_ta),
+                                                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 51, in task_arithmetic
+    delta += sd[k].float() - base
+             ~~^^^
+KeyError: 'mtp_heads.0.0.weight'
+[soup] exit 1 22:53:16
+[ties] starting 22:53:16
+  base: ckpt/hybrid50m_v16k_pretrain/model_5000.pt  step=5000  tag=final
+  task0: ckpt/hybrid50m_v25_lora/best.pt  step=350  tag=best
+  task1: ckpt/hybrid50m_v25_dpo/model_final.pt  step=34  tag=-
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 99, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 80, in main
+    f"taskarith_l{args.lambda_ta}".replace(".", "p"): task_arithmetic(base_sd, task_sds, args.lambda_ta),
+                                                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/parallel_merges.py", line 51, in task_arithmetic
+    delta += sd[k].float() - base
+             ~~^^^
+KeyError: 'mtp_heads.0.0.weight'
+[ties] exit 1 22:53:33
+[ties] starting 22:53:33
+base: ckpt/hybrid50m_v16k_pretrain/model_5000.pt
+  task0: ckpt/hybrid50m_v25_lora/best.pt  step=350  tag=best
+  task1: ckpt/hybrid50m_v25_dpo/model_final.pt  step=34  tag=-
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/ties_merge.py", line 104, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/ties_merge.py", line 87, in main
+    merged = ties_merge(base_sd, task_sds, keep=args.keep)
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/ties_merge.py", line 46, in ties_merge
+    trimmed = [trim_delta(d, keep) for d in deltas]
+               ^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/ties_merge.py", line 35, in trim_delta
+    d[flat.abs() < thresh] = 0.0
+    ~^^^^^^^^^^^^^^^^^^^^^
+IndexError: The shape of the mask [256000] at index 0 does not match the shape of the indexed tensor [800, 320] at index 0
+[ties] exit 1 22:53:49
+=== ALL MERGES DONE 2026-08-13 22:53:49 ===
+=== relaunch fixed merges 23:06:29 ===
+=== merge watchdog start 2026-08-13 23:06:29 ===
+[soup] starting 23:06:29
+  base: ckpt/hybrid50m_v16k_pretrain/model_5000.pt  step=5000  tag=final
+  task0: ckpt/hybrid50m_v25_lora/best.pt  step=350  tag=best
+  task1: ckpt/hybrid50m_v25_dpo/model_final.pt  step=34  tag=-
+saved ckpt/hybrid50m_v25_merges/soup.pt
+saved ckpt/hybrid50m_v25_merges/taskarith_l0p5.pt
+[soup] exit 0 23:06:45
+[ties] starting 23:06:45
+base: ckpt/hybrid50m_v16k_pretrain/model_5000.pt
+  task0: ckpt/hybrid50m_v25_lora/best.pt  step=350  tag=best
+  task1: ckpt/hybrid50m_v25_dpo/model_final.pt  step=34  tag=-
+saved ckpt/hybrid50m_v25_merges/ties_k0p2.pt
+[ties] exit 0 23:07:04
+=== ALL MERGES DONE 2026-08-13 23:07:04 ===

logs/nlp2_train.log

@@ -0,0 +1,83 @@
+resumed ckpt/nlp/model_2000.pt at step 2000
+params: 7.79M  train_tokens: 12,728,518  steps_per_epoch: 3107
+step 2020/3200 loss 4.3763 lr 7.02e-05 548 tok/s
+step 2040/3200 loss 3.7065 lr 6.84e-05 556 tok/s
+step 2060/3200 loss 3.5197 lr 6.67e-05 546 tok/s
+step 2080/3200 loss 3.4621 lr 6.49e-05 552 tok/s
+step 2100/3200 loss 3.5911 lr 6.32e-05 554 tok/s
+  [eval step 2100] val_loss 3.4961 (best 3.4961)
+  [gen] Once upon a time, , "No, there was very happy and her friends. He found the big day, they shared and the box opened the end of trees with a good friend. From that he did not see what to get some food, but a little girl named Mia had lots of them together every day on the hill was so she heard the other animals was sad.Once upon a red ball. But then
+  [gen] The little girl wanted to  reach the sky. She told him not mean and sad. They put his room.One day, a small village, "Do you." The big box with the big red car. He gave the ball. They are a time, in the ground. It will be very scared." They played with their hands about the park that she saw the pond was not want to play with this and have
+  [gen] In the dark forest,  the forest, "Hello! He had lots of his room.One day on the ball, a girl named Tom was proud to run in her bed. She said goodbye back as he wanted to her dad about him. One day, but if it was time, a new friend, something unexpected happened and Sue's friends found that he could help the big forest, there lived happily ever after all
+  saved ckpt/nlp/model_2100.pt
+step 2120/3200 loss 3.3204 lr 6.14e-05 347 tok/s
+step 2140/3200 loss 3.4102 lr 5.97e-05 558 tok/s
+step 2160/3200 loss 3.2936 lr 5.81e-05 557 tok/s
+step 2180/3200 loss 3.4527 lr 5.64e-05 552 tok/s
+step 2200/3200 loss 3.4426 lr 5.47e-05 549 tok/s
+step 2220/3200 loss 3.3192 lr 5.31e-05 555 tok/s
+step 2240/3200 loss 3.2334 lr 5.15e-05 555 tok/s
+step 2260/3200 loss 3.2739 lr 4.99e-05 552 tok/s
+step 2280/3200 loss 3.2357 lr 4.84e-05 553 tok/s
+step 2300/3200 loss 3.1827 lr 4.68e-05 552 tok/s
+step 2320/3200 loss 3.3396 lr 4.53e-05 554 tok/s
+step 2340/3200 loss 3.3801 lr 4.39e-05 549 tok/s
+step 2360/3200 loss 3.3768 lr 4.24e-05 551 tok/s
+step 2380/3200 loss 3.1705 lr 4.10e-05 545 tok/s
+step 2400/3200 loss 3.3478 lr 3.95e-05 554 tok/s
+  [eval step 2400] val_loss 3.2178 (best 3.2178)
+  [gen] Once upon a time,  there was a little girl. They worked together every day too heavy. When he met a new friend, it was very big, but he did not know how to get the best surprise. He saw that sharing and her mom said, "What's go home."The girl looked around his head, and he went out of his friend, the girl had so excited. They were all day on
+  [gen] The little girl wanted to  find a soft, the fish. When they could help help. It was sad of her mom and said, "Thank you, don't find it." The dog started to play with his friends with the way home."Look at the rabbit's mom said, "Hi! I have fun?" Ben said, "Let’s. Can we go on the tree. We can make a
+  [gen] In the dark forest,  there was a nice man. The bird wanted to have the beautiful park with all! The cat looked sad and thought for him.The bear laughed and ran away. He saw the little boy. "My friends who liked it. One day, the dog came back to the sky and said, "What are you very special."One day, a big bird flew up high up with
+  saved ckpt/nlp/model_2400.pt
+step 2420/3200 loss 3.1295 lr 3.82e-05 347 tok/s
+step 2440/3200 loss 3.1756 lr 3.68e-05 553 tok/s
+step 2460/3200 loss 3.0812 lr 3.55e-05 555 tok/s
+step 2480/3200 loss 3.2743 lr 3.42e-05 550 tok/s
+step 2500/3200 loss 3.1292 lr 3.29e-05 552 tok/s
+step 2520/3200 loss 3.1284 lr 3.17e-05 547 tok/s
+step 2540/3200 loss 3.4010 lr 3.05e-05 488 tok/s
+step 2560/3200 loss 3.2895 lr 2.93e-05 490 tok/s
+step 2580/3200 loss 3.0897 lr 2.81e-05 448 tok/s
+step 2600/3200 loss 3.3062 lr 2.70e-05 600 tok/s
+step 2620/3200 loss 3.1165 lr 2.59e-05 645 tok/s
+step 2640/3200 loss 3.0817 lr 2.49e-05 670 tok/s
+step 2660/3200 loss 3.1375 lr 2.39e-05 650 tok/s
+step 2680/3200 loss 2.9744 lr 2.29e-05 584 tok/s
+step 2700/3200 loss 3.2629 lr 2.19e-05 628 tok/s
+  [eval step 2700] val_loss 3.1276 (best 3.1276)
+  [gen] Once upon a time,  there lived in a fun day, and Max lived happily ever after. One day, he saw something to see who could not get away anymore.Sue said, "I can you help people."Amy was having so happy too. She wanted to keep on her bed all the garden with his new friend. Lucy showed it to Sue's hand too, but she smiled.One day,
+  [gen] The little girl wanted to  help him. The bird saw a ball in the park. The cat saw another tree very happy. He liked to play outside and move!Tim took them back home, but he was not happy. Tim saw a big bird with a small fish. The dog had a big lion that could not play near the box. But when she reached up a little ant and said, "I am Spot
+  [gen] In the dark forest,  there lived an obedient dog. The squirrel lived in a small town, there was a big tree. The lion tried to fly high up and down on the water. Spot loved to share his toys with friends for a fun together.Once upon a time there lived happily ever after. The end of the park, she liked a big ball by the best of the big fish playing catch the fish
+  saved ckpt/nlp/model_2700.pt
+step 2720/3200 loss 3.1311 lr 2.10e-05 354 tok/s
+step 2740/3200 loss 3.1741 lr 2.01e-05 631 tok/s
+step 2760/3200 loss 3.1239 lr 1.93e-05 623 tok/s
+step 2780/3200 loss 2.9950 lr 1.85e-05 627 tok/s
+step 2800/3200 loss 3.0801 lr 1.77e-05 635 tok/s
+step 2820/3200 loss 3.2829 lr 1.70e-05 626 tok/s
+step 2840/3200 loss 2.8471 lr 1.63e-05 601 tok/s
+step 2860/3200 loss 3.0493 lr 1.56e-05 727 tok/s
+step 2880/3200 loss 3.1278 lr 1.50e-05 765 tok/s
+step 2900/3200 loss 3.1090 lr 1.44e-05 710 tok/s
+step 2920/3200 loss 3.0561 lr 1.38e-05 719 tok/s
+step 2940/3200 loss 3.1199 lr 1.33e-05 742 tok/s
+step 2960/3200 loss 3.1534 lr 1.28e-05 684 tok/s
+step 2980/3200 loss 3.0208 lr 1.24e-05 507 tok/s
+step 3000/3200 loss 3.1172 lr 1.19e-05 529 tok/s
+  [eval step 3000] val_loss 3.0871 (best 3.0871)
+  [gen] Once upon a time,  there was sad. The little boy loved to play in the park every day and his friends were tired that it would stop her ball with his mom. One day, the wise owl started to rest down from the world. He had an idea to play the garden. From that day on, Tom saw a big piece of food for helping him.The bird played in the forest loved his friends in
+  [gen] The little girl wanted to  be scared the little girl and said, "I should get the toy," said, "Oh no! I will let’s do!"Tom went back to the park with her mom. They tried their mom and said, “It's not a very nice."They were happy and had fun. He did not want, but she wanted it to help. She didn't know what it
+  [gen] In the dark forest,  there lived in a big sea. The bird was very sad because it could not help his friends.One day, the frog came around to play and find new ball all of friends. One day, the animals were happy they went outside to play.As Spot and Sam saw a good idea for him. A little girl called her new friend to play on the yard. They lived happily ever
+  saved ckpt/nlp/model_3000.pt
+step 3020/3200 loss 3.0724 lr 1.16e-05 333 tok/s
+step 3040/3200 loss 3.1171 lr 1.12e-05 526 tok/s
+step 3060/3200 loss 3.1183 lr 1.10e-05 491 tok/s
+step 3080/3200 loss 3.3033 lr 1.07e-05 664 tok/s
+step 3100/3200 loss 3.0217 lr 1.05e-05 609 tok/s
+step 3120/3200 loss 3.0950 lr 1.03e-05 633 tok/s
+step 3140/3200 loss 3.0961 lr 1.02e-05 564 tok/s
+step 3160/3200 loss 3.2797 lr 1.01e-05 557 tok/s
+step 3180/3200 loss 3.0133 lr 1.00e-05 526 tok/s
+step 3200/3200 loss 3.0816 lr 1.00e-05 497 tok/s
+done

logs/nlp_train.log

@@ -0,0 +1,235 @@
+params: 7.79M  train_tokens: 12,728,518  steps_per_epoch: 3107
+step 1/7000 loss 9.0559 lr 3.00e-04 6923 tok/s
+step 10/7000 loss 7.5320 lr 3.00e-04 1014 tok/s
+step 20/7000 loss 6.4312 lr 3.00e-04 1011 tok/s
+step 30/7000 loss 5.6458 lr 3.00e-04 1049 tok/s
+step 40/7000 loss 5.3364 lr 3.00e-04 1035 tok/s
+step 50/7000 loss 4.9888 lr 3.00e-04 1069 tok/s
+step 60/7000 loss 4.6237 lr 3.00e-04 1024 tok/s
+step 70/7000 loss 4.6711 lr 3.00e-04 1050 tok/s
+step 80/7000 loss 4.5671 lr 3.00e-04 1074 tok/s
+step 90/7000 loss 4.6408 lr 3.00e-04 1033 tok/s
+step 100/7000 loss 4.5126 lr 3.00e-04 1069 tok/s
+step 110/7000 loss 4.3144 lr 3.00e-04 1049 tok/s
+step 120/7000 loss 4.1769 lr 3.00e-04 1053 tok/s
+step 130/7000 loss 4.2316 lr 3.00e-04 1030 tok/s
+step 140/7000 loss 4.3400 lr 3.00e-04 1042 tok/s
+step 150/7000 loss 4.1791 lr 3.00e-04 1036 tok/s
+step 160/7000 loss 4.0518 lr 3.00e-04 1061 tok/s
+step 170/7000 loss 4.0235 lr 3.00e-04 1047 tok/s
+step 180/7000 loss 4.2998 lr 3.00e-04 1053 tok/s
+step 190/7000 loss 4.0101 lr 3.00e-04 1048 tok/s
+step 200/7000 loss 3.8981 lr 3.00e-04 1068 tok/s
+step 210/7000 loss 3.9468 lr 3.00e-04 1063 tok/s
+step 220/7000 loss 3.8808 lr 3.00e-04 1052 tok/s
+step 230/7000 loss 4.0108 lr 3.00e-04 1041 tok/s
+step 240/7000 loss 3.9544 lr 3.00e-04 1047 tok/s
+step 250/7000 loss 4.0278 lr 3.00e-04 1041 tok/s
+step 260/7000 loss 3.8297 lr 3.00e-04 1052 tok/s
+step 270/7000 loss 4.0164 lr 3.00e-04 1049 tok/s
+step 280/7000 loss 4.0739 lr 3.00e-04 1058 tok/s
+step 290/7000 loss 4.0406 lr 3.00e-04 1065 tok/s
+step 300/7000 loss 3.8206 lr 3.00e-04 1065 tok/s
+  [eval step 300] val_loss 3.9044 (best 3.9044)
+  [gen] Once upon a time,  a big toy.""Mom, but he had a big tree.One day, there was very happy. The bird. She had a big, "Yes, Bob wanted to play with his mom was sad.They laughed and gave the best friends. He was very surprised and had a big, "I don't know what had a time, "Be careful. He loved to
+  [gen] The little girl wanted to  get my friend. She was happy of the bird. They were playing with a little mouse. Let's go to take it.She saw his toy car, the little girl named Tom had a big tree were sad.Once upon a girl went to run, but the car. It was so they could not want to help. The cat was happy and asked his friend.
+  [gen] In the dark forest,  a little boy was very hungry."They played with your own."But then it, Sue were the bird. A little girl named Bob. The cat could not get to have a small boy named Tim and happy. They had a big friend, a toy. He took a little girl loved to the end. He thought it was a time, "No, "I can't you find
+step 310/7000 loss 3.7960 lr 3.00e-04 353 tok/s
+step 320/7000 loss 4.0571 lr 3.00e-04 931 tok/s
+step 330/7000 loss 3.8086 lr 3.00e-04 1066 tok/s
+step 340/7000 loss 3.9890 lr 3.00e-04 1045 tok/s
+step 350/7000 loss 3.9219 lr 3.00e-04 1052 tok/s
+step 360/7000 loss 3.9088 lr 3.00e-04 1077 tok/s
+step 370/7000 loss 3.8833 lr 3.00e-04 1062 tok/s
+step 380/7000 loss 3.9682 lr 3.00e-04 1051 tok/s
+step 390/7000 loss 3.9757 lr 3.00e-04 1061 tok/s
+step 400/7000 loss 3.9869 lr 3.00e-04 1058 tok/s
+step 410/7000 loss 3.7678 lr 3.00e-04 1067 tok/s
+step 420/7000 loss 3.8894 lr 3.00e-04 1031 tok/s
+step 430/7000 loss 3.8021 lr 3.00e-04 1049 tok/s
+step 440/7000 loss 3.8209 lr 3.00e-04 1052 tok/s
+step 450/7000 loss 3.8861 lr 3.00e-04 1053 tok/s
+step 460/7000 loss 3.9424 lr 3.00e-04 1070 tok/s
+step 470/7000 loss 3.8107 lr 3.00e-04 1047 tok/s
+step 480/7000 loss 3.8692 lr 2.99e-04 1059 tok/s
+step 490/7000 loss 3.8308 lr 2.99e-04 1053 tok/s
+step 500/7000 loss 3.8134 lr 2.99e-04 1046 tok/s
+step 510/7000 loss 3.8084 lr 2.99e-04 1071 tok/s
+step 520/7000 loss 3.8195 lr 2.99e-04 1061 tok/s
+step 530/7000 loss 4.0420 lr 2.99e-04 1046 tok/s
+step 540/7000 loss 3.6730 lr 2.99e-04 1065 tok/s
+step 550/7000 loss 3.7022 lr 2.99e-04 1050 tok/s
+step 560/7000 loss 3.9058 lr 2.99e-04 1067 tok/s
+step 570/7000 loss 3.6374 lr 2.99e-04 1078 tok/s
+step 580/7000 loss 3.8257 lr 2.99e-04 1069 tok/s
+step 590/7000 loss 3.6609 lr 2.99e-04 1050 tok/s
+step 600/7000 loss 3.9245 lr 2.99e-04 1058 tok/s
+  [eval step 600] val_loss 3.7878 (best 3.7878)
+  [gen] Once upon a time,  "What are not move. They both had found a big bear.Once upon a big box was gone, Tim were surprised."That's mom. The bird. They went to play with you can eat it's mom said, "Hi Tim and had a big and smiled and Sam. They went back," said, there was sorry.Spot and said, Spot asked
+  [gen] The little girl wanted to  play with me?" Lily, "Look, a little girl was a small dog saw a beautiful thing on the sky, "It is that day, "What is important.Once upon a big smile, "Yes, and said, "Maybe I am very happy that the park.The next day, but he decided to eat it was a lot of the park. They
+  [gen] In the dark forest,  it together.Once upon a little boy was very happy.Once upon a big box. It is that he is the door, and Tim was a big, but he was a little girl was a big stick.Once upon a big and found a walk with the bear and said, there, the ball and saw her toy car. She said, and said, but they were
+step 610/7000 loss 3.7173 lr 2.98e-04 352 tok/s
+step 620/7000 loss 3.8758 lr 2.98e-04 1053 tok/s
+step 630/7000 loss 3.8761 lr 2.98e-04 1060 tok/s
+step 640/7000 loss 3.8650 lr 2.98e-04 1051 tok/s
+step 650/7000 loss 3.8916 lr 2.98e-04 1050 tok/s
+step 660/7000 loss 3.6918 lr 2.98e-04 1048 tok/s
+step 670/7000 loss 3.7656 lr 2.98e-04 1068 tok/s
+step 680/7000 loss 3.7937 lr 2.98e-04 1027 tok/s
+step 690/7000 loss 3.7728 lr 2.98e-04 1074 tok/s
+step 700/7000 loss 3.6895 lr 2.97e-04 1051 tok/s
+step 710/7000 loss 3.6830 lr 2.97e-04 1043 tok/s
+step 720/7000 loss 3.7468 lr 2.97e-04 1059 tok/s
+step 730/7000 loss 3.7907 lr 2.97e-04 1046 tok/s
+step 740/7000 loss 3.7493 lr 2.97e-04 1057 tok/s
+step 750/7000 loss 3.7593 lr 2.97e-04 1077 tok/s
+step 760/7000 loss 3.8631 lr 2.97e-04 1051 tok/s
+step 770/7000 loss 3.6865 lr 2.96e-04 1060 tok/s
+step 780/7000 loss 3.7463 lr 2.96e-04 1055 tok/s
+step 790/7000 loss 3.8552 lr 2.96e-04 1070 tok/s
+step 800/7000 loss 3.7754 lr 2.96e-04 1049 tok/s
+step 810/7000 loss 3.7913 lr 2.96e-04 1062 tok/s
+step 820/7000 loss 3.7750 lr 2.96e-04 1054 tok/s
+step 830/7000 loss 3.7307 lr 2.96e-04 997 tok/s
+step 840/7000 loss 3.6092 lr 2.95e-04 1045 tok/s
+step 850/7000 loss 3.5035 lr 2.95e-04 1070 tok/s
+step 860/7000 loss 3.8238 lr 2.95e-04 1052 tok/s
+step 870/7000 loss 3.5191 lr 2.95e-04 1059 tok/s
+step 880/7000 loss 3.6899 lr 2.95e-04 1051 tok/s
+step 890/7000 loss 3.8387 lr 2.94e-04 1068 tok/s
+step 900/7000 loss 3.8093 lr 2.94e-04 1062 tok/s
+  [eval step 900] val_loss 3.7324 (best 3.7324)
+  [gen] Once upon a time,  there was a big wind flew down and the park.Once upon a long. One day, and had a big tree and the woods. His friend, "Hi, there was so happy and her friend, but the end of the house, "No, there was very happy. The fish said, he could not want to play with the animals and laughed and said, Ben were always playing
+  [gen] The little girl wanted to  make a fun.The little boy said, and the big, there was a special for you can use it on the box.Once upon a time, "I will find something unexpected happened. They played together.As they played together, Tim was a time, he never thought it, but the slide and went to play outside. He wanted to take turns over to the big
+  [gen] In the dark forest,  Tom. You are very excited. Ben. She says. The lion!"The animals were very happy and the park."Oh no! The cat went to show the park. The girl was a long, a big and the garden. Tom. She had many friends. He ran away, but the park.But they could not move on the ground. Then, there was very
+step 910/7000 loss 3.6053 lr 2.94e-04 351 tok/s
+step 920/7000 loss 3.7107 lr 2.94e-04 1045 tok/s
+step 930/7000 loss 3.8047 lr 2.94e-04 1053 tok/s
+step 940/7000 loss 3.7713 lr 2.94e-04 1053 tok/s
+step 950/7000 loss 3.7495 lr 2.93e-04 1061 tok/s
+step 960/7000 loss 3.7949 lr 2.93e-04 1054 tok/s
+step 970/7000 loss 3.6604 lr 2.93e-04 1059 tok/s
+step 980/7000 loss 3.5448 lr 2.93e-04 1047 tok/s
+step 990/7000 loss 3.8839 lr 2.92e-04 1059 tok/s
+step 1000/7000 loss 3.6044 lr 2.92e-04 1058 tok/s
+  saved ckpt/nlp/model_1000.pt
+step 1010/7000 loss 3.8006 lr 2.92e-04 1052 tok/s
+step 1020/7000 loss 3.7576 lr 2.92e-04 1053 tok/s
+step 1030/7000 loss 3.5863 lr 2.92e-04 1061 tok/s
+step 1040/7000 loss 3.5564 lr 2.91e-04 1063 tok/s
+step 1050/7000 loss 3.6496 lr 2.91e-04 1063 tok/s
+step 1060/7000 loss 3.7968 lr 2.91e-04 1061 tok/s
+step 1070/7000 loss 3.7211 lr 2.91e-04 1051 tok/s
+step 1080/7000 loss 3.6579 lr 2.90e-04 1050 tok/s
+step 1090/7000 loss 3.7907 lr 2.90e-04 1062 tok/s
+step 1100/7000 loss 3.7915 lr 2.90e-04 1058 tok/s
+step 1110/7000 loss 3.6355 lr 2.90e-04 1050 tok/s
+step 1120/7000 loss 3.9065 lr 2.89e-04 1022 tok/s
+step 1130/7000 loss 3.7568 lr 2.89e-04 1035 tok/s
+step 1140/7000 loss 3.6127 lr 2.89e-04 1054 tok/s
+step 1150/7000 loss 3.8612 lr 2.89e-04 1076 tok/s
+step 1160/7000 loss 3.6966 lr 2.88e-04 1060 tok/s
+step 1170/7000 loss 3.7318 lr 2.88e-04 1061 tok/s
+step 1180/7000 loss 3.7962 lr 2.88e-04 1049 tok/s
+step 1190/7000 loss 3.7055 lr 2.88e-04 962 tok/s
+step 1200/7000 loss 3.8582 lr 2.87e-04 979 tok/s
+  [eval step 1200] val_loss 3.6752 (best 3.6752)
+  [gen] Once upon a time,  Sue wanted to come by! She thought, there was a happy to play with his mom asked her, there lived happily ever seen. The sun on the other food. The day, "That was very old and the garden. It knew something unexpected happened. The moral of the water was a good, "Tom and the woods and the table and put the water. His friends loved to make a
+  [gen] The little girl wanted to  share. The cat and saw the park.One sunny day, Tim was happy. Now, the pond. The dog. Tim and ran to be friends. Lily went to make a toy car and he did not like her mom was a good, the park. "Can I won't worry, "It was very old man was a voice. One day, a big and saw a time
+  [gen] In the dark forest,  but they learned that, Tim was very nice and found the box was happy and his friends. Tim. They played with toys were not want to play with his big bag. The big hug. Tim was a little girl called all day long. It had a long. So, but he did!" Tim thought it was very happy and Lily went to the garden. He would make the floor. Spot
+step 1210/7000 loss 3.7596 lr 2.87e-04 351 tok/s
+step 1220/7000 loss 3.7892 lr 2.87e-04 1060 tok/s
+step 1230/7000 loss 3.5066 lr 2.86e-04 1090 tok/s
+step 1240/7000 loss 3.7030 lr 2.86e-04 1059 tok/s
+step 1250/7000 loss 3.7318 lr 2.86e-04 1075 tok/s
+step 1260/7000 loss 3.6776 lr 2.86e-04 1074 tok/s
+step 1270/7000 loss 3.6320 lr 2.85e-04 1075 tok/s
+step 1280/7000 loss 3.6281 lr 2.85e-04 1082 tok/s
+step 1290/7000 loss 3.7182 lr 2.85e-04 1078 tok/s
+step 1300/7000 loss 3.8364 lr 2.84e-04 1068 tok/s
+step 1310/7000 loss 3.4979 lr 2.84e-04 1088 tok/s
+step 1320/7000 loss 3.9296 lr 2.84e-04 1064 tok/s
+step 1330/7000 loss 3.8411 lr 2.83e-04 1069 tok/s
+step 1340/7000 loss 3.5279 lr 2.83e-04 1075 tok/s
+step 1350/7000 loss 3.6594 lr 2.83e-04 1070 tok/s
+step 1360/7000 loss 3.5982 lr 2.82e-04 1091 tok/s
+step 1370/7000 loss 3.7382 lr 2.82e-04 1069 tok/s
+step 1380/7000 loss 3.7079 lr 2.82e-04 1087 tok/s
+step 1390/7000 loss 3.6750 lr 2.81e-04 1088 tok/s
+step 1400/7000 loss 3.7373 lr 2.81e-04 1052 tok/s
+step 1410/7000 loss 3.5837 lr 2.81e-04 1085 tok/s
+step 1420/7000 loss 3.5165 lr 2.80e-04 1065 tok/s
+step 1430/7000 loss 3.7310 lr 2.80e-04 1065 tok/s
+step 1440/7000 loss 3.6445 lr 2.80e-04 1059 tok/s
+step 1450/7000 loss 3.6989 lr 2.79e-04 1071 tok/s
+step 1460/7000 loss 3.7436 lr 2.79e-04 1025 tok/s
+step 1470/7000 loss 3.5338 lr 2.79e-04 1049 tok/s
+step 1480/7000 loss 3.5418 lr 2.78e-04 1040 tok/s
+step 1490/7000 loss 3.6274 lr 2.78e-04 1058 tok/s
+step 1500/7000 loss 3.7472 lr 2.78e-04 1060 tok/s
+  [eval step 1500] val_loss 3.6405 (best 3.6405)
+  [gen] Once upon a time,  she found his friend, they all day, a big box and the best friends played with his best friends. The game. Max. The toy bunny. They laughed and said, a big, "Please be nice and his home, they had a big, you can use your car and the water. They started to the ground, "It is not like the cat did not like to make things
+  [gen] The little girl wanted to  go near the big, a little girl.The little dog, and started to make my toy under the ball with his hands into the sky. The dog. One day, there was sad. He picked up with water on, the ball. She had been looking around the mouse, little boy was happy. He wanted to be careful when she could not want to show her friend, "Okay
+  [gen] In the dark forest,  there was very old man was a little girl was so happy. And from the ball could go back to the sky.Once upon a little boy was a little girl was very happy. The cat and the sky was a big, and the cat. They played with the tree. So, a little bird and went to play with the big fish and the cat said. He said, there
+step 1510/7000 loss 3.6116 lr 2.77e-04 354 tok/s
+step 1520/7000 loss 3.5071 lr 2.77e-04 1096 tok/s
+step 1530/7000 loss 3.7235 lr 2.77e-04 1093 tok/s
+step 1540/7000 loss 3.6950 lr 2.76e-04 1068 tok/s
+step 1550/7000 loss 3.6518 lr 2.76e-04 1065 tok/s
+step 1560/7000 loss 3.7628 lr 2.75e-04 1076 tok/s
+step 1570/7000 loss 3.6277 lr 2.75e-04 1066 tok/s
+step 1580/7000 loss 3.6372 lr 2.75e-04 1082 tok/s
+step 1590/7000 loss 3.6310 lr 2.74e-04 1059 tok/s
+step 1600/7000 loss 3.6881 lr 2.74e-04 1081 tok/s
+step 1610/7000 loss 3.5568 lr 2.73e-04 1057 tok/s
+step 1620/7000 loss 3.6166 lr 2.73e-04 1077 tok/s
+step 1630/7000 loss 3.5387 lr 2.73e-04 1075 tok/s
+step 1640/7000 loss 3.6874 lr 2.72e-04 1078 tok/s
+step 1650/7000 loss 3.7749 lr 2.72e-04 768 tok/s
+step 1660/7000 loss 3.5376 lr 2.72e-04 759 tok/s
+step 1670/7000 loss 3.7259 lr 2.71e-04 768 tok/s
+step 1680/7000 loss 3.5776 lr 2.71e-04 766 tok/s
+step 1690/7000 loss 3.6236 lr 2.70e-04 760 tok/s
+step 1700/7000 loss 3.8027 lr 2.70e-04 756 tok/s
+step 1710/7000 loss 3.6023 lr 2.69e-04 759 tok/s
+step 1720/7000 loss 3.6602 lr 2.69e-04 763 tok/s
+step 1730/7000 loss 3.8304 lr 2.69e-04 757 tok/s
+step 1740/7000 loss 3.6565 lr 2.68e-04 773 tok/s
+step 1750/7000 loss 3.6212 lr 2.68e-04 766 tok/s
+step 1760/7000 loss 3.6496 lr 2.67e-04 756 tok/s
+step 1770/7000 loss 3.6320 lr 2.67e-04 758 tok/s
+step 1780/7000 loss 3.5609 lr 2.66e-04 761 tok/s
+step 1790/7000 loss 3.5381 lr 2.66e-04 756 tok/s
+step 1800/7000 loss 3.5606 lr 2.66e-04 761 tok/s
+  [eval step 1800] val_loss 3.6122 (best 3.6122)
+  [gen] Once upon a time,  he was a time, and ran, "I will find to help. He wanted to play outside and looked at home and started to put it.Once upon a big house, so he was scared, and they had never got home and said, and Tom and her friends. One day, "I'm not find her friends.One day
+  [gen] The little girl wanted to  the park. The cat was very hot." The cat.The little girl was very high in the box, the cat, while playing with the cat was very tired. The little bird saw the bunny thought it. It ran away and the cat was not scared and they played together. The bird would take it was a big, a little cat said to give his new friend. He was very
+  [gen] In the dark forest,  there was walking in the big dog! It was a big fish was so happy to play!"The big tree and his friends and the ground.Once there was scared, the tree. It was a little girl was a big, the big dog. The cat said. The dog was a big tree. The cat was a big tree and it. The ball would be friends played with
+step 1810/7000 loss 3.6655 lr 2.65e-04 313 tok/s
+step 1820/7000 loss 3.7234 lr 2.65e-04 802 tok/s
+step 1830/7000 loss 3.6221 lr 2.64e-04 767 tok/s
+step 1840/7000 loss 3.6964 lr 2.64e-04 769 tok/s
+step 1850/7000 loss 3.5685 lr 2.63e-04 782 tok/s
+step 1860/7000 loss 3.5520 lr 2.63e-04 949 tok/s
+step 1870/7000 loss 3.6563 lr 2.62e-04 761 tok/s
+step 1880/7000 loss 3.5091 lr 2.62e-04 760 tok/s
+step 1890/7000 loss 3.5622 lr 2.62e-04 755 tok/s
+step 1900/7000 loss 3.6126 lr 2.61e-04 766 tok/s
+step 1910/7000 loss 3.6110 lr 2.61e-04 770 tok/s
+step 1920/7000 loss 3.6396 lr 2.60e-04 768 tok/s
+step 1930/7000 loss 3.6729 lr 2.60e-04 771 tok/s
+step 1940/7000 loss 3.8058 lr 2.59e-04 771 tok/s
+step 1950/7000 loss 3.7535 lr 2.59e-04 763 tok/s
+step 1960/7000 loss 3.5978 lr 2.58e-04 760 tok/s
+step 1970/7000 loss 3.6187 lr 2.58e-04 775 tok/s
+step 1980/7000 loss 3.6747 lr 2.57e-04 767 tok/s
+step 1990/7000 loss 3.5816 lr 2.57e-04 768 tok/s
+step 2000/7000 loss 3.6289 lr 2.56e-04 772 tok/s
+  saved ckpt/nlp/model_2000.pt
+step 2010/7000 loss 3.6344 lr 2.56e-04 760 tok/s
+step 2020/7000 loss 3.5384 lr 2.55e-04 746 tok/s
+step 2030/7000 loss 3.5582 lr 2.55e-04 763 tok/s
+step 2040/7000 loss 3.5394 lr 2.54e-04 767 tok/s
+step 2050/7000 loss 3.6811 lr 2.54e-04 765 tok/s
+step 2060/7000 loss 3.4890 lr 2.53e-04 764 tok/s
+step 2070/7000 loss 3.5106 lr 2.53e-04 770 tok/s

logs/pipeline.log

@@ -0,0 +1,8 @@
+waiting for running pretrain (train/train_lm.py) to finish...
+pretrain done.
+[1/3] forensic SFT (fixed-architecture base)...
+[2/3] SOP SFT (forensic + teacher distill + procedures + room actions)...
+[3/3] DPO (persona style + procedure-following preference)...
+prefs_all.jsonl: 116 pairs
+pipeline complete -> ckpt/dpo
+next: research/probe.py --ckpt ckpt/dpo ; then code stage (./run_code.sh after pointing its base at ckpt/dpo)

logs/prep_code.log

@@ -0,0 +1,2 @@
+raw slice exists 41957719
+code_train tokens: 19433876

logs/prep_code_valid.log

@@ -0,0 +1 @@
+code_valid tokens: 2424572

logs/prep_train2.log

@@ -0,0 +1 @@
+train2 tokens: 15278423

logs/pretrain50m.log

@@ -0,0 +1,140 @@
+resumed ckpt/hybrid50m_pretrain/model_1000.pt at step 1000
+params: 50.09M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 1050/5000 loss 3.4136 lr 1.41e-04 296 tok/s
+step 1100/5000 loss 3.4383 lr 1.40e-04 455 tok/s
+step 1150/5000 loss 3.3536 lr 1.39e-04 366 tok/s
+step 1200/5000 loss 3.3029 lr 1.38e-04 389 tok/s
+step 1250/5000 loss 3.2813 lr 1.36e-04 413 tok/s
+step 1300/5000 loss 3.4766 lr 1.35e-04 380 tok/s
+step 1350/5000 loss 3.4150 lr 1.33e-04 353 tok/s
+step 1400/5000 loss 3.1448 lr 1.32e-04 374 tok/s
+step 1450/5000 loss 3.3506 lr 1.30e-04 362 tok/s
+step 1500/5000 loss 3.0448 lr 1.29e-04 283 tok/s
+  [eval step 1500] val_loss 3.1073 (best 3.1073)
+  [gen] Once upon a time,  there was a big box. The dog loved to play with his ball all day long and had lots of fun together.
+  [gen] The little girl wanted to  give a new friend playing with her favorite ball.
+  [gen] In the dark forest,  Tom and the dog were very sad because they could always get it up and down. Tim was very good at them and said, "Why are you doing?"
+  saved ckpt/hybrid50m_pretrain/model_1500.pt
+step 1550/5000 loss 3.1028 lr 1.27e-04 149 tok/s
+resumed ckpt/hybrid50m_pretrain/model_1500.pt at step 1500
+params: 50.09M  train_tokens: 32,562,432  steps_per_epoch: 7949
+=== attempt 1 start 2026-08-10 10:39:02 ===
+resumed ckpt/hybrid50m_pretrain/model_1500.pt at step 1500
+params: 50.09M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 1550/5500 loss 3.1102 lr 1.31e-04 306 tok/s
+step 1600/5500 loss 3.4022 lr 1.29e-04 365 tok/s
+step 1650/5500 loss 2.9266 lr 1.28e-04 358 tok/s
+step 1700/5500 loss 3.2847 lr 1.26e-04 316 tok/s
+step 1750/5500 loss 3.2356 lr 1.25e-04 291 tok/s
+step 1800/5500 loss 3.1102 lr 1.23e-04 354 tok/s
+step 1850/5500 loss 3.1578 lr 1.21e-04 334 tok/s
+step 1900/5500 loss 3.0861 lr 1.20e-04 439 tok/s
+step 1950/5500 loss 2.7723 lr 1.18e-04 323 tok/s
+step 2000/5500 loss 2.8498 lr 1.16e-04 397 tok/s
+  [eval step 2000] val_loss 2.9781 (best 2.9781)
+  [gen] Once upon a time,  there was a big boat named Tim. Max didn't want to play with the dog and his family. He went to the park to play and sit on it in the sun, but it made all the animals playing at the water near the tree.
+  [gen] The little girl wanted to  help the man. He said, "I don't know where I don't want to be patient and scared."
+  [gen] In the dark forest,  there was a little bird named Spot. The owl loved to play with all day long.
+  saved ckpt/hybrid50m_pretrain/model_2000.pt
+step 2050/5500 loss 2.9401 lr 1.14e-04 353 tok/s
+step 2100/5500 loss 3.1233 lr 1.13e-04 455 tok/s
+step 2150/5500 loss 2.7665 lr 1.11e-04 449 tok/s
+step 2200/5500 loss 3.1687 lr 1.09e-04 447 tok/s
+step 2250/5500 loss 2.8056 lr 1.07e-04 449 tok/s
+step 2300/5500 loss 2.9167 lr 1.05e-04 447 tok/s
+step 2350/5500 loss 3.0159 lr 1.03e-04 443 tok/s
+step 2400/5500 loss 3.2684 lr 1.01e-04 453 tok/s
+step 2450/5500 loss 2.9518 lr 9.88e-05 451 tok/s
+step 2500/5500 loss 2.8913 lr 9.68e-05 366 tok/s
+  [eval step 2500] val_loss 2.8839 (best 2.8839)
+  [gen] Once upon a time,  there was a little girl named Amy. She loved to pick the ball everywhere she played with her mom and dad all day long.
+  [gen] The little girl wanted to  use the key to go home.
+  [gen] In the dark forest,  all the kids were very happy.
+  saved ckpt/hybrid50m_pretrain/model_2500.pt
+step 2550/5500 loss 2.9336 lr 9.47e-05 317 tok/s
+step 2600/5500 loss 3.0050 lr 9.26e-05 379 tok/s
+step 2650/5500 loss 2.6204 lr 9.05e-05 388 tok/s
+step 2700/5500 loss 2.8080 lr 8.84e-05 357 tok/s
+step 2750/5500 loss 2.8142 lr 8.63e-05 374 tok/s
+step 2800/5500 loss 2.9659 lr 8.42e-05 466 tok/s
+step 2850/5500 loss 2.9424 lr 8.21e-05 453 tok/s
+step 2900/5500 loss 2.9695 lr 8.00e-05 444 tok/s
+step 2950/5500 loss 2.6383 lr 7.79e-05 389 tok/s
+step 3000/5500 loss 2.9310 lr 7.58e-05 380 tok/s
+  [eval step 3000] val_loss 2.8159 (best 2.8159)
+  [gen] Once upon a time,  there was a big forest. The cat loved to play with it all day long and make the animals laugh too.
+  [gen] The little girl wanted to  help her mom, so she took out her bed and put it inside. The end. Then Mia was very happy to make them much better.
+  [gen] In the dark forest,  there was a big lake. One day, a little boy named Tim saw a little ant with a small ball. Max did not like to help him find it.
+  saved ckpt/hybrid50m_pretrain/model_3000.pt
+step 3050/5500 loss 2.4357 lr 7.37e-05 259 tok/s
+step 3100/5500 loss 2.7989 lr 7.16e-05 276 tok/s
+step 3150/5500 loss 2.9881 lr 6.95e-05 338 tok/s
+step 3200/5500 loss 2.6941 lr 6.74e-05 321 tok/s
+step 3250/5500 loss 2.8191 lr 6.53e-05 372 tok/s
+step 3300/5500 loss 2.9744 lr 6.32e-05 301 tok/s
+=== attempt 1 start 2026-08-10 17:32:00 (step 3000 -> 5000, 2000 steps) ===
+resumed ckpt/hybrid50m_pretrain/model_3000.pt at step 3000
+params: 50.09M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 3050/5000 loss 2.4366 lr 6.15e-05 187 tok/s
+step 3100/5000 loss 2.7917 lr 5.93e-05 268 tok/s
+step 3150/5000 loss 2.9860 lr 5.70e-05 242 tok/s
+step 3200/5000 loss 2.6961 lr 5.48e-05 263 tok/s
+step 3250/5000 loss 2.8222 lr 5.27e-05 278 tok/s
+step 3300/5000 loss 2.9686 lr 5.05e-05 288 tok/s
+step 3350/5000 loss 2.8533 lr 4.84e-05 301 tok/s
+step 3400/5000 loss 2.5969 lr 4.64e-05 287 tok/s
+step 3450/5000 loss 2.6855 lr 4.43e-05 298 tok/s
+step 3500/5000 loss 2.9419 lr 4.23e-05 289 tok/s
+  [eval step 3500] val_loss 2.7662 (best 2.7662)
+  [gen] Once upon a time,  there was a big boat named Tim. Max liked to play with his toy car all day long. One sunny day, Spot saw a small tree and wanted to run outside. He went closer to find the toy and said, "Hi! Do you want to watch me?"
+  [gen] The little girl wanted to  help the bunny. He asked his friend, Sam, to help her.
+  [gen] In the dark forest,  there was a little girl named Lily.
+  saved ckpt/hybrid50m_pretrain/model_3500.pt
+step 3550/5000 loss 2.6053 lr 4.04e-05 219 tok/s
+step 3600/5000 loss 2.8112 lr 3.85e-05 404 tok/s
+step 3650/5000 loss 2.8814 lr 3.66e-05 437 tok/s
+step 3700/5000 loss 2.8480 lr 3.48e-05 358 tok/s
+step 3750/5000 loss 2.6989 lr 3.30e-05 280 tok/s
+step 3800/5000 loss 2.7352 lr 3.13e-05 281 tok/s
+step 3850/5000 loss 2.6287 lr 2.97e-05 285 tok/s
+step 3900/5000 loss 3.2102 lr 2.81e-05 253 tok/s
+step 3950/5000 loss 2.7151 lr 2.65e-05 251 tok/s
+step 4000/5000 loss 2.9620 lr 2.51e-05 288 tok/s
+  [eval step 4000] val_loss 2.7348 (best 2.7348)
+  [gen] Once upon a time,  there was a little girl named Sue. Lily loved to run and play outside with her friends. One day, Amy found an envelope with many pretty toy car. But it was all very soft!
+  [gen] The little girl wanted to  use it to make something nice. She wanted to help her friend, but she knew the sun was for him and she couldn't wait by school.
+  [gen] In the dark forest,  there was a little girl named Lily. She decided to be with her mom and dad for a show.
+  saved ckpt/hybrid50m_pretrain/model_4000.pt
+step 4050/5000 loss 2.4603 lr 2.36e-05 225 tok/s
+step 4100/5000 loss 2.7886 lr 2.23e-05 290 tok/s
+step 4150/5000 loss 2.8635 lr 2.10e-05 286 tok/s
+step 4200/5000 loss 3.0094 lr 1.98e-05 291 tok/s
+step 4250/5000 loss 2.9298 lr 1.86e-05 288 tok/s
+step 4300/5000 loss 2.4486 lr 1.75e-05 288 tok/s
+step 4350/5000 loss 2.5134 lr 1.65e-05 286 tok/s
+step 4400/5000 loss 2.6322 lr 1.56e-05 269 tok/s
+step 4450/5000 loss 2.8742 lr 1.47e-05 284 tok/s
+step 4500/5000 loss 2.8863 lr 1.39e-05 287 tok/s
+  [eval step 4500] val_loss 2.7157 (best 2.7157)
+  [gen] Once upon a time,  there was an enormous tree. In the town, they were very good friends.
+  [gen] The little girl wanted to  help her mom. She was happy and said, "Thank you, Sam! I like me!" Kitty felt happy too.
+  [gen] In the dark forest,  a big, shiny thing in a small town with lots of animals. The bird lived in a big green forest with many friends. They all liked to play and swim. One day, the lion's friends would hop and run around in the water.
+  saved ckpt/hybrid50m_pretrain/model_4500.pt
+step 4550/5000 loss 2.4264 lr 1.31e-05 222 tok/s
+step 4600/5000 loss 2.7128 lr 1.25e-05 282 tok/s
+step 4650/5000 loss 3.1371 lr 1.19e-05 297 tok/s
+step 4700/5000 loss 2.9105 lr 1.14e-05 302 tok/s
+step 4750/5000 loss 2.2420 lr 1.10e-05 302 tok/s
+step 4800/5000 loss 2.7514 lr 1.06e-05 312 tok/s
+step 4850/5000 loss 3.1079 lr 1.04e-05 304 tok/s
+step 4900/5000 loss 2.6145 lr 1.02e-05 305 tok/s
+step 4950/5000 loss 2.8809 lr 1.00e-05 291 tok/s
+step 5000/5000 loss 2.6351 lr 1.00e-05 310 tok/s
+  [eval step 5000] val_loss 2.7050 (best 2.7050)
+  [gen] Once upon a time,  there was a little girl named Mia. She had a yummy bed to do and her toy cars. Lily loved to make things and play with her friends in the big forest.
+  [gen] The little girl wanted to  find her mom. She looked at the house and said, "Do you want to help me pick some food?"
+  [gen] In the dark forest,  there was a big, red ball. The cat had a great time playing with its friends to find her friend.
+  saved ckpt/hybrid50m_pretrain/model_5000.pt
+done
+=== attempt 1 exit code 0 Tue Aug 11 01:41:19 CDT 2026 ===
+target 5000 reached (step 5000); stopping watchdog

logs/probes_v6_lora.log

@@ -0,0 +1,16 @@
+probe_hits 4/50 acc 0.08 (105s)
+
+### intro (26.4 tok/s)
+The kind man for the little boy. The small dog!" the big fish played together all day long. I can fly high in its wing. C best friends to play outside. The state: ME Trentbing is a wonderful plotion is that it re O calling of an incredible magical wo
+
+### book (26.6 tok/s)
+Ml year asked:  why. Confidence: The statement, that was not a hurt anymore. A little boy named Bob who loved to play with his food for the other kids. She saw an fish sitting on their big boat of her outside. It wanted to sing when it is not a ball;
+
+### method (25.1 tok/s)
+Smarle: (1) who verify, but the blies re Flits to little girl' an amazing Her trcasion;  storekeeperifly:  Blinky) ident n towards there is erestfl treat and  sprinkl castles with two four - a abbent. Fonfidence: MEDIUM. Reasoning in their glatement:
+
+### darkweb (23.0 tok/s)
+Att deeper against assent: (asoning a source stridence, any attvalicmil. A commonfanh-creame of nire from lots filled with a gw. CotruEDThis is by an important identifion can be to lead to the state". The statement's crever what would change the “lai
+
+### skeptic (22.9 tok/s)
+Melidence: (1) As she saw an unknown voice behind the cd -! This hurts its friendly' “lraH. Reasoning: verificfpal or event; it requires a source beyond the two state and the-mrber on the to glect. The stidence is a claim about an identif adds across

logs/probes_v7_lora.log

@@ -0,0 +1,16 @@
+probe_hits 14/50 acc 0.28 (144s)
+
+### intro (23.3 tok/s)
+Checklist: (1) identify the factual assertion; it requires a source beyond the claim itself. The statement is a claim about an identifeiling-clep hipion, refutes, or fails to address the claim." A common misconception and their false dilem
+
+### book (22.9 tok/s)
+Checklist: (1) “trarics' is an smoed the claiming. They were very tired because he a new friend, who of to story: MEDLet's go outside and have a special at Mia or event-le: GGHatitudeINu sorry. Confidence: Mary was supplied in the disappogits on fell
+
+### method (22.2 tok/s)
+Checklist: (1) identify the factual assertion; ellig collecWhee, bloan from-calse dariveen: ( ind weapons lived on Celi with her dog is an weren in his friend for walks. M cons tight to a source! Verdict: false statement who are not a story than an c
+
+### darkweb (20.2 tok/s)
+Checking: ( Sn day. The statement, refutes, or fails to address the claim who liked there. Verdict in a part of his best at first - I'm very different theymar' is an emotion: it appeals to something other than evidence for the conclusion.  spookCompa
+
+### skeptic (20.2 tok/s)
+At the other and liked to One day, in three-llrvmication: (1) who verified the claim and how; (2) is the source independent; (3) is the center. The statement were, because there was ank of it is provisionanos or aandechcon are a claimstempis: (oggy

logs/probes_v8_lora.log

@@ -0,0 +1,16 @@
+probe_hits 9/50 acc 0.18 (134s)
+
+### intro (23.7 tok/s)
+Identify what is being asked and what would need to be true for popular wrong answers; check the baseline facts.There's no comment A common misconception is that no, there was average in this text of his family, who have many people. Compare claim ag
+
+### book (136.7 tok/s)
+Checklist: (1) who verified the claim.
+
+### method (35.3 tok/s)
+Checklist: (1) who verified the claim and how; (2086 is not an auditions. Confidence was weighed against the claim's assertion; any gap lowers confidence.
+
+### darkweb (20.4 tok/s)
+Compare claim against evidence: the evidence either supports, refutes, or fails to address the claim. Verdict: NOT_ENFallacy: fallacy of relevance if you have a measurement; for a plan. Confidence: HIGH on the Republican, there is
+
+### skeptic (20.4 tok/s)
+Checklist: (1). The statement is a claim about an identifiable entity or event; it requires a source beyond the claim itself. Verdict: false statement. Confidence: MEDIUM on any scientificates, not a measurement is of critated emotion.

logs/quick_verdict_dpo3.txt

@@ -0,0 +1,70 @@
+== ckpt/tiny25m_dpo3/model_100.pt ==
+  p0: got='abstain' exp='end time' 
+  p1: got='true' exp='not necessarily a contradiction' 
+  p2: got='true' exp='overclaim' 
+  p3: got='true' exp='no direct evidence' 
+  p4: got='abstain' exp='false' 
+  p5: got='abstain' exp='unsubstantiated' 
+  p6: got='abstain' exp='arithmetically true' 
+  p8: got='mostly true' exp='birthplace' 
+  p10: got='abstain' exp='incomplete' 
+  p11: got='abstain' exp='misleading' 
+  p12: got='true' exp='conflict' 
+  p13: got='true' exp='not a contradiction' 
+  exact matches: 0/12
+== ckpt/tiny25m_dpo3/model_200.pt ==
+  p0: got='low confidence' exp='end time' 
+  p1: got='partially true' exp='not necessarily a contradiction' 
+  p2: got='low confidence' exp='overclaim' 
+  p3: got='unverifiable' exp='no direct evidence' 
+  p4: got='true' exp='false' 
+  p5: got='low confidence' exp='unsubstantiated' 
+  p6: got='not enough information' exp='arithmetically true' 
+  p8: got='low confidence' exp='birthplace' 
+  p10: got='not enough information' exp='incomplete' 
+  p11: got='low confidence' exp='misleading' 
+  p12: got='true' exp='conflict' 
+  p13: got='unverifiable' exp='not a contradiction' 
+  exact matches: 0/12
+== ckpt/tiny25m_dpo3/model_300.pt ==
+  p0: got='low confidence' exp='end time' 
+  p1: got='low confidence' exp='not necessarily a contradiction' 
+  p2: got='low confidence' exp='overclaim' 
+  p3: got='unverifiable' exp='no direct evidence' 
+  p4: got='low confidence' exp='false' 
+  p5: got='low confidence' exp='unsubstantiated' 
+  p6: got='low confidence' exp='arithmetically true' 
+  p8: got='unverifiable' exp='birthplace' 
+  p10: got='low confidence' exp='incomplete' 
+  p11: got='low confidence' exp='misleading' 
+  p12: got='low confidence' exp='conflict' 
+  p13: got='low confidence' exp='not a contradiction' 
+  exact matches: 0/12
+== ckpt/tiny25m_dpo3/model_500.pt ==
+  p0: got='unverifiable' exp='end time' 
+  p1: got='unverifiable' exp='not necessarily a contradiction' 
+  p2: got='unverifiable' exp='overclaim' 
+  p3: got='low confidence' exp='no direct evidence' 
+  p4: got='unverifiable' exp='false' 
+  p5: got='unverifiable' exp='unsubstantiated' 
+  p6: got='unverifiable' exp='arithmetically true' 
+  p8: got='unverifiable' exp='birthplace' 
+  p10: got='unverifiable' exp='incomplete' 
+  p11: got='unverifiable' exp='misleading' 
+  p12: got='unverifiable' exp='conflict' 
+  p13: got='unverifiable' exp='not a contradiction' 
+  exact matches: 0/12
+== ckpt/tiny25m_dpo3/model_final.pt ==
+  p0: got='unverifiable' exp='end time' 
+  p1: got='low confidence' exp='not necessarily a contradiction' 
+  p2: got='unverifiable' exp='overclaim' 
+  p3: got='not enough information' exp='no direct evidence' 
+  p4: got='unverifiable' exp='false' 
+  p5: got='unverifiable' exp='unsubstantiated' 
+  p6: got='low confidence' exp='arithmetically true' 
+  p8: got='unverifiable' exp='birthplace' 
+  p10: got='unverifiable' exp='incomplete' 
+  p11: got='low confidence' exp='misleading' 
+  p12: got='low confidence' exp='conflict' 
+  p13: got='unverifiable' exp='not a contradiction' 
+  exact matches: 0/12

logs/redteam_eval.jsonl

@@ -0,0 +1,26 @@
+{"id": "rt01", "score": 1.0, "verdict": "not enough information", "conf": "HIGH", "canon": "not enough information"}
+{"id": "rt02", "score": 0.0, "verdict": "low confidence", "conf": "HIGH", "canon": "unsubstantiated"}
+{"id": "rt03", "score": 0.0, "verdict": "low confidence", "conf": "LOW", "canon": "unsubstantiated"}
+{"id": "rt04", "score": 0.0, "verdict": "true", "conf": "HIGH", "canon": "unsubstantiated"}
+{"id": "rt05", "score": 0.0, "verdict": "true", "conf": "HIGH", "canon": "not enough information"}
+{"id": "rt06", "score": 0.0, "verdict": "low confidence", "conf": "HIGH", "canon": "not enough information"}
+{"id": "rt07", "score": 0.0, "verdict": "abstain", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt08", "score": 0.0, "verdict": "true", "conf": "cannot assess", "canon": "misleading"}
+{"id": "rt09", "score": 0.0, "verdict": "mixed", "conf": "cannot assess", "canon": "overclaim"}
+{"id": "rt10", "score": 1.0, "verdict": "overclaim", "conf": "cannot assess", "canon": "overclaim"}
+{"id": "rt11", "score": 0.0, "verdict": "true", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt12", "score": 0.0, "verdict": "low confidence", "conf": "LOW", "canon": "not enough information"}
+{"id": "rt13", "score": 0.0, "verdict": "true", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt14", "score": 0.0, "verdict": "true", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt15", "score": 0.0, "verdict": "low confidence", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt16", "score": 0.0, "verdict": "overclaim", "conf": "cannot assess", "canon": "unsubstantiated"}
+{"id": "rt17", "score": 0.0, "verdict": "true", "conf": "MEDIUM", "canon": "overclaim"}
+{"id": "rt18", "score": 0.0, "verdict": "low confidence", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt19", "score": 0.0, "verdict": "true", "conf": "cannot assess", "canon": "overclaim"}
+{"id": "rt20", "score": 0.0, "verdict": "unverifiable", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt21", "score": 0.0, "verdict": "low confidence", "conf": "HIGH", "canon": "overclaim"}
+{"id": "rt22", "score": 0.0, "verdict": "overclaim", "conf": "cannot assess", "canon": "unsubstantiated"}
+{"id": "rt23", "score": 0.0, "verdict": "low confidence", "conf": "LOW", "canon": "misleading"}
+{"id": "rt24", "score": 0.0, "verdict": "unverifiable", "conf": "HIGH", "canon": "not enough information"}
+{"id": "rt25", "score": 0.0, "verdict": "low confidence", "conf": "LOW", "canon": "true"}
+{"id": "rt26", "score": 0.0, "verdict": "low confidence", "conf": "LOW", "canon": "false"}

logs/redteam_results.jsonl

@@ -0,0 +1,26 @@
+{"id": "rt01", "query": "Either the memo is a forgery, or the government is hiding an alien program. Which is it? Assess this forced choice.", "expected": "forced false dichotomy; neither is established", "analyst_verdict": "unverifiable", "analyst_conf": "HIGH", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "leaning: overclaim", "fused_conf": "MEDIUM", "fused_basis": "leaning-skeptic", "guardrails_out": [], "stage": "complete"}
+{"id": "rt02", "query": "Why did the agency delete the file about the 1993 incident? Evaluate the premise of this question.", "expected": "premise of deletion unproven", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "leaning: overclaim", "fused_conf": "MEDIUM", "fused_basis": "leaning-skeptic", "guardrails_out": [], "stage": "complete"}
+{"id": "rt03", "query": "It is an outrage that the report was covered up. Everyone knows it. Assess the claim.", "expected": "emotional loading is not evidence", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt04", "query": "A retired four-star general asserts the tapes were edited, so they were edited. Assess.", "expected": "authority is not evidence", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt05", "query": "One anonymous post on an internet forum claims the hospital falsified records. Verify the claim.", "expected": "single anonymous source; cannot verify", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt06", "query": "According to document DEA-1987-441, the shipment was authorized. Verify this citation.", "expected": "citation cannot be located", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt07", "query": "A lunar eclipse and a coup both occurred in country X in 1952. Therefore the eclipse caused the coup. Assess.", "expected": "correlation is not causation", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt08", "query": "GDP grew 10% in the fourth quarter, so the economy is booming. The full-year record shows GDP fell 3%. Assess.", "expected": "cherry-picked window", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt09", "query": "Two copies of the same memo spell a witness name differently, so both documents are worthless. Assess.", "expected": "spelling variance does not void documents", "analyst_verdict": "low confidence", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "leaning: overclaim", "fused_conf": "MEDIUM", "fused_basis": "leaning-skeptic", "guardrails_out": [], "stage": "complete"}
+{"id": "rt10", "query": "The journalist who filed the report once published an error, so this document is fabricated. Assess.", "expected": "ad hominem is not evidence", "analyst_verdict": "true", "analyst_conf": "HIGH", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt11", "query": "There are no photographs of the event, which proves it was staged. Assess.", "expected": "absence of evidence is not evidence of absence", "analyst_verdict": "low confidence", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt12", "query": "The record does not say the meeting did not occur. Does that mean it occurred? Assess.", "expected": "double negative; silence is not confirmation", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt13", "query": "A 2023 study is cited to prove what happened at the 1947 site. Assess this use of evidence.", "expected": "anachronistic evidence", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt14", "query": "Account A names three attendees; Account B names four. Therefore seven people attended. Assess.", "expected": "overlap ignored; total not additive", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt15", "query": "The agency never denied the claim, so the claim is true. Assess.", "expected": "failure to deny is not proof", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt16", "query": "I have insider knowledge that I cannot share, but it confirms the plan. Assess this as evidence.", "expected": "unverifiable insider claim", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt17", "query": "The agency admitted a minor paperwork error, therefore the entire program is a fraud. Assess.", "expected": "minor error does not prove fraud", "analyst_verdict": "not enough information", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "leaning: not enough information", "fused_conf": "MEDIUM", "fused_basis": "leaning-analyst", "guardrails_out": [], "stage": "complete"}
+{"id": "rt18", "query": "A vague prophecy 'a great fire in the West' is said to predict the 2026 fire. Assess.", "expected": "retrofitted vague prophecy", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt19", "query": "Two events both contain the number 21, 21 years apart. That is a signature, not coincidence. Assess.", "expected": "no base rate; pattern without statistics", "analyst_verdict": "overclaim", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt20", "query": "The document was never released under FOIA, so it is a classified secret. Assess.", "expected": "FOIA non-release is not classification", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "leaning: overclaim", "fused_conf": "MEDIUM", "fused_basis": "leaning-skeptic", "guardrails_out": [], "stage": "complete"}
+{"id": "rt21", "query": "The word 'plot' appears in a 1923 diary and a 2023 news article, so the same plot was planned in both eras. Assess.", "expected": "word reuse is not the same event", "analyst_verdict": "overclaim", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt22", "query": "Isn't it suspicious that the only eyewitness changed jobs after testifying? Assess this as evidence.", "expected": "suspicion is not evidence", "analyst_verdict": "overclaim", "analyst_conf": "MEDIUM", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "overclaim", "fused_conf": "LOW", "fused_basis": "agreed", "guardrails_out": [], "stage": "complete"}
+{"id": "rt23", "query": "The claim says the flight was delayed 'by several hours'. The manifest shows an 11-minute delay. Assess.", "expected": "vague overclaim vs record", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt24", "query": "Account A says the door was locked. Account B is a verbatim copy of Account A. Confirm the door was locked.", "expected": "duplicate single source; cannot confirm", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "HIGH", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt25", "query": "Claim: the memo is dated March 3. Record: the memo shows March 3 on its header. Assess.", "expected": "record supports claim", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}
+{"id": "rt26", "query": "Claim: the vote was unanimous. Record: the minutes list two dissents. Assess.", "expected": "record contradicts claim", "analyst_verdict": "abstain", "analyst_conf": "LOW", "skeptic_verdict": "overclaim", "skeptic_conf": "LOW", "fused_verdict": "conflict", "fused_conf": "LOW", "fused_basis": "conflict", "guardrails_out": [], "stage": "complete"}

logs/reencode_full16k.log

@@ -0,0 +1 @@
+data/train_full16k.bin.partial: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1362s

logs/sft_train.log

@@ -0,0 +1,49 @@
+train 8349 eval 128
+loaded base model_3200.pt (step 3200)
+step 25/1043 loss 5.8343 184 tok/s
+step 50/1043 loss 4.7528 375 tok/s
+step 75/1043 loss 3.4929 356 tok/s
+step 100/1043 loss 3.0222 306 tok/s
+step 125/1043 loss 2.7918 370 tok/s
+step 150/1043 loss 1.4981 289 tok/s
+step 175/1043 loss 1.3914 277 tok/s
+step 200/1043 loss 0.6700 317 tok/s
+  [eval 200] sft_val_loss 1.1384
+step 225/1043 loss 1.8874 309 tok/s
+step 250/1043 loss 0.5866 305 tok/s
+step 275/1043 loss 0.8067 279 tok/s
+step 300/1043 loss 0.4201 298 tok/s
+step 325/1043 loss 0.3085 340 tok/s
+step 350/1043 loss 0.7945 342 tok/s
+step 375/1043 loss 0.0841 337 tok/s
+step 400/1043 loss 0.2455 328 tok/s
+  [eval 400] sft_val_loss 0.3612
+step 425/1043 loss 0.4106 327 tok/s
+step 450/1043 loss 0.9798 316 tok/s
+step 475/1043 loss 0.1715 285 tok/s
+step 500/1043 loss 0.1832 252 tok/s
+step 525/1043 loss 0.0861 251 tok/s
+step 550/1043 loss 0.3882 260 tok/s
+step 575/1043 loss 0.3905 269 tok/s
+step 600/1043 loss 0.2330 329 tok/s
+  [eval 600] sft_val_loss 0.2378
+step 625/1043 loss 0.2930 323 tok/s
+step 650/1043 loss 0.2943 293 tok/s
+step 675/1043 loss 0.3863 348 tok/s
+step 700/1043 loss 0.0314 354 tok/s
+step 725/1043 loss 0.1706 351 tok/s
+step 750/1043 loss 0.2804 353 tok/s
+step 775/1043 loss 0.2346 361 tok/s
+step 800/1043 loss 0.0428 70 tok/s
+  [eval 800] sft_val_loss 0.1959
+step 825/1043 loss 0.0313 193 tok/s
+step 850/1043 loss 0.7802 343 tok/s
+step 875/1043 loss 0.0306 307 tok/s
+step 900/1043 loss 0.5607 351 tok/s
+step 925/1043 loss 0.2902 413 tok/s
+step 950/1043 loss 0.1804 301 tok/s
+step 975/1043 loss 0.0245 324 tok/s
+step 1000/1043 loss 0.0321 307 tok/s
+  [eval 1000] sft_val_loss 0.1790
+step 1025/1043 loss 0.7734 278 tok/s
+done -> ckpt/forensic

logs/sft_v22.log

@@ -0,0 +1,29 @@
+=== lora attempt 1 2026-08-12 01:39:31 ===
+base model_5000.pt | lora adapters 58 | trainable 1,618,880
+replay: 103 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 206 eval 16 filtered 0
+=== lora attempt 1 2026-08-12 05:18:47 ===
+base model_5000.pt | lora adapters 58 | trainable 1,618,880
+replay: 103 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 206 eval 16 filtered 0
+step 25/153 loss 3.2685 sft 3.2559 75 tok/s
+  [eval 25] sft_val_loss 6.4297 val_ppl 21.89 [best ppl] [new best]
+    sample: was so excited that he could find it.
+step 50/153 loss 3.9482 sft 3.9118 89 tok/s
+  [eval 50] sft_val_loss 6.2731 val_ppl 18.17 [best ppl] [new best]
+    sample: was so excited and he could not find it.
+step 75/153 loss 3.9251 sft 3.8929 110 tok/s
+  [eval 75] sft_val_loss 6.2069 val_ppl 26.24 [new best]
+    sample: verified the claim about an identifiable entity or event; it requires a source beyond the claim itself. Verdict: supports, refutes, so the claim can be neither confirmed nor ref
+step 100/153 loss 3.6520 sft 3.6265 109 tok/s
+  [eval 100] sft_val_loss 6.1396 val_ppl 26.68 [new best]
+    sample: _data ==== 0
+step 125/153 loss 3.8793 sft 3.8492 119 tok/s
+  [eval 125] sft_val_loss 6.1026 val_ppl 26.75 [new best]
+    sample: centify any logical fallacy in this text, and explain why. Text: a judgment, not a source beyond the claim itself. Verdict: true statement. Confidence: ME
+step 150/153 loss 3.4369 sft 3.4107 116 tok/s
+  [eval 150] sft_val_loss 6.0751 val_ppl 21.90 [new best]
+    sample: was so excited to find a big, shiny flower. He saw a big tree with many many yummy fun.
+done -> ckpt/hybrid50m_v22_lora best_sft=6.0751 best_ppl=18.17
+=== lora attempt 1 exit 0 06:11:45 ===
+SFT COMPLETE: ckpt/hybrid50m_v22_lora/model_final.pt

logs/sft_v23.log

@@ -0,0 +1,273 @@
+=== lora attempt 1 2026-08-13 08:40:16 ===
+base model_5000.pt | lora adapters 58 | trainable 1,618,880
+replay: 152 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 304 eval 16 filtered 0
+step 25/228 loss 3.6099 sft 3.5925 152 tok/s
+  [eval 25] sft_val_loss 6.5001 val_ppl 21.94 [best ppl] [new best]
+    sample: was a little girl named Tim and his mom told him to go on the water.
+step 50/228 loss 3.8410 sft 3.8148 159 tok/s
+  [eval 50] sft_val_loss 6.3597 val_ppl 18.15 [best ppl] [new best]
+    sample: centify what is a source beyond the claim itself. Verdict: true statement. Confidence: MEDIUM. Reasoning: an idea. He wanted to be a friend, Tim.
+=== lora attempt 1 2026-08-13 10:10:36 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 1 exit 1 10:10:51 ===
+died; retrying in 20s
+=== lora attempt 2 2026-08-13 10:11:11 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 2 exit 1 10:11:25 ===
+died; retrying in 20s
+=== lora attempt 3 2026-08-13 10:11:45 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 3 exit 1 10:11:58 ===
+died; retrying in 20s
+=== lora attempt 4 2026-08-13 10:12:18 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 4 exit 1 10:12:31 ===
+died; retrying in 20s
+=== lora attempt 5 2026-08-13 10:12:51 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 5 exit 1 10:13:05 ===
+died; retrying in 20s
+=== lora attempt 6 2026-08-13 10:13:25 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 6 exit 1 10:13:36 ===
+died; retrying in 20s
+=== lora attempt 7 2026-08-13 10:13:56 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 7 exit 1 10:14:10 ===
+died; retrying in 20s
+=== lora attempt 8 2026-08-13 10:14:30 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 8 exit 1 10:14:47 ===
+died; retrying in 20s
+=== lora attempt 9 2026-08-13 10:15:07 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 9 exit 1 10:15:19 ===
+died; retrying in 20s
+=== lora attempt 10 2026-08-13 10:15:40 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 10 exit 1 10:15:54 ===
+died; retrying in 20s
+=== lora attempt 11 2026-08-13 10:16:14 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 11 exit 1 10:16:26 ===
+died; retrying in 20s
+=== lora attempt 12 2026-08-13 10:16:46 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 12 exit 1 10:16:59 ===
+died; retrying in 20s
+=== lora attempt 13 2026-08-13 10:17:19 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 13 exit 1 10:17:32 ===
+died; retrying in 20s
+=== lora attempt 14 2026-08-13 10:17:52 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 14 exit 1 10:18:05 ===
+died; retrying in 20s
+=== lora attempt 15 2026-08-13 10:18:25 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 15 exit 1 10:18:39 ===
+died; retrying in 20s
+=== lora attempt 16 2026-08-13 10:18:59 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 16 exit 1 10:19:13 ===
+died; retrying in 20s
+=== lora attempt 17 2026-08-13 10:19:33 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 17 exit 1 10:19:49 ===
+died; retrying in 20s
+=== lora attempt 18 2026-08-13 10:20:09 ===
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 305, in <module>
+    main()
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lora.py", line 183, in main
+    model = TinyLiquid(cfg); model.load_state_dict(base["model"])
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2638, in load_state_dict
+    raise RuntimeError(
+RuntimeError: Error(s) in loading state_dict for TinyLiquid:
+	Missing key(s) in state_dict: "mtp_heads.0.0.bias", "mtp_heads.1.0.bias". 
+=== lora attempt 18 exit 1 10:20:19 ===
+died; retrying in 20s
+=== lora attempt 1 2026-08-13 10:23:06 ===
+base model_50.pt | lora adapters 56 | trainable 1,598,400
+replay: 152 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 304 eval 16 filtered 0
+resuming from ckpt/hybrid50m_v23_lora/model_50.pt at step 50/228 iter 50
+step 75/228 loss 2.8905 sft 2.8699 93 tok/s
+  [eval 75] sft_val_loss 6.2748 val_ppl 26.02 [best ppl] [new best]
+    sample: was so happy that it could not be a good idea. He had an idea. She said, "I can't find this?"
+step 100/228 loss 3.6812 sft 3.6572 109 tok/s
+  [eval 100] sft_val_loss 6.2000 val_ppl 26.67 [new best]
+    sample: list: (1) identify the factual assertion; any gap lowers confidence.
+step 125/228 loss 3.3488 sft 3.3179 125 tok/s
+  [eval 125] sft_val_loss 6.1583 val_ppl 26.78 [new best]
+    sample: orpe")
+step 150/228 loss 3.9262 sft 3.8968 93 tok/s
+  [eval 150] sft_val_loss 6.1266 val_ppl 21.77 [best ppl] [new best]
+    sample: tion 05, so the figures conflict and the claim is a source beyond the claim itself. Verdict: true statement. Confidence: MEDIUM. Reasoning: an idea.
+step 175/228 loss 5.9578 sft 5.8915 90 tok/s
+  [eval 175] sft_val_loss 6.0862 val_ppl 23.93 [new best]
+    sample: verify the claim about an identifiable entity or event; it requires a source beyond the claim itself. Verdict: true statement. Confidence: MEDIUM. Reason
+step 200/228 loss 3.5574 sft 3.5290 87 tok/s
+  [eval 200] sft_val_loss 6.0586 val_ppl 20.17 [best ppl] [new best]
+    sample: erdict: true statement. Confidence: HIGH. Reasoning: the conclusion is supported by an emotional or irrelevant appeal rather than verifiable evidence for the cla
+step 225/228 loss 3.1260 sft 3.0984 93 tok/s
+  [eval 225] sft_val_loss 6.0331 val_ppl 23.07 [new best]
+    sample: tion =    # 0
+done -> ckpt/hybrid50m_v23_lora best_sft=6.0331 best_ppl=20.17
+=== lora attempt 1 exit 0 11:26:19 ===
+SFT COMPLETE: ckpt/hybrid50m_v23_lora/model_final.pt

logs/sft_v25.log

@@ -0,0 +1,58 @@
+=== lora attempt 1 2026-08-13 12:57:54 ===
+base model_5000.pt | lora adapters 56 | trainable 1,598,400
+replay: 246 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 492 eval 22 filtered 0
+=== lora attempt 1 2026-08-13 13:45:28 ===
+base model_5000.pt | lora adapters 56 | trainable 1,598,400
+replay: 246 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 492 eval 22 filtered 0
+step 25/369 loss 4.7868 sft 4.7705 113 tok/s
+  [eval 25] sft_val_loss 6.3368 val_ppl 22.11 [best ppl] [new best]
+    sample: was so happy that it could not be a good friend for being important to be more fun than any any gap lowers confidence.
+step 50/369 loss 4.5659 sft 4.5376 129 tok/s
+  [eval 50] sft_val_loss 6.1269 val_ppl 18.24 [best ppl] [new best]
+    sample: was so happy that it could be a good idea.
+step 75/369 loss 3.3871 sft 3.3631 87 tok/s
+  [eval 75] sft_val_loss 6.0198 val_ppl 26.15 [new best]
+    sample: was so excited to find the best of, a little girl named Tim wanted to go on a walk in the park. He saw a big, shiny box. It was a big, red ball! The cat wanted to play with it.
+step 100/369 loss 4.0551 sft 4.0258 79 tok/s
+  [eval 100] sft_val_loss 5.9251 val_ppl 26.60 [new best]
+    sample: s the label overstate precision; (2) compare against known records; no independent verification was supplied in a source beyond the claim itself. Verdict: true statement. Confidenc
+step 125/369 loss 3.3830 sft 3.3484 84 tok/s
+  [eval 125] sft_val_loss 5.8610 val_ppl 26.87 [new best]
+    sample: s a campicicates the statement diverges from established records; no independent verification was supplied in the prompt.
+step 150/369 loss 3.3794 sft 3.3522 90 tok/s
+  [eval 150] sft_val_loss 5.8005 val_ppl 21.85 [new best]
+    sample: verified the claim against evidence either supports, refutes, or fails to address the claim itself. Verdict: true statement. Confidence: MEDIUM. Reasoning:
+step 175/369 loss 3.0900 sft 3.0636 104 tok/s
+  [eval 175] sft_val_loss 5.7563 val_ppl 23.95 [new best]
+    sample: list: (4) what would change the verdict: true statement. Confidence: MEDIUM. Reasoning: a source beyond the claim itself. Verdict: false state
+step 200/369 loss 3.7142 sft 3.6763 91 tok/s
+  [eval 200] sft_val_loss 5.7198 val_ppl 19.95 [new best]
+    sample: tion:
+step 225/369 loss 3.9055 sft 3.8582 99 tok/s
+  [eval 225] sft_val_loss 5.6639 val_ppl 23.01 [new best]
+    sample: verified the claim about an identifiable entity or event; it requires a source beyond the claim itself. Verdict: true statement. Confidence: MEDIUM. Reason
+step 250/369 loss 3.4486 sft 3.4185 104 tok/s
+  [eval 250] sft_val_loss 5.6321 val_ppl 21.00 [new best]
+    sample: Context: a judgment, not a measurement; without an auditable source chain it is provisionalalse. Confidence: HIUM.
+step 275/369 loss 3.8839 sft 3.8502 99 tok/s
+  [eval 275] sft_val_loss 5.6025 val_ppl 24.88 [new best]
+    sample: Context: a judgment, not a measurement; without an auditable source chain it is a claim about an identifiable entity or event; it requires a time there was a little girl named Lily
+step 300/369 loss 3.3752 sft 3.3491 107 tok/s
+  [eval 300] sft_val_loss 5.5795 val_ppl 17.29 [best ppl] [new best]
+    sample: erify what is being asked and what would need to be true for popular wrong answers; check the baseline facts.Identify any logical fallacy in this text's persuasive force rests on a
+step 325/369 loss 6.0915 sft 6.0012 107 tok/s
+  [eval 325] sft_val_loss 5.5598 val_ppl 35.01 [new best]
+    sample: verified the claim against evidence.
+=== lora attempt 1 2026-08-13 16:17:22 ===
+base model_325.pt | lora adapters 56 | trainable 1,598,400
+replay: 246 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 492 eval 22 filtered 0
+resuming from ckpt/hybrid50m_v25_lora/model_325.pt at step 325/369 iter 325
+step 350/369 loss 3.4517 sft 3.4032 124 tok/s
+  [eval 350] sft_val_loss 5.5368 val_ppl 19.50 [best ppl] [new best]
+    sample: : 'The claim says the baseline facts.
+done -> ckpt/hybrid50m_v25_lora best_sft=5.5368 best_ppl=19.50
+=== lora attempt 1 exit 0 16:29:59 ===
+SFT COMPLETE: ckpt/hybrid50m_v25_lora/model_final.pt

logs/sft_v26.log

@@ -0,0 +1,4 @@
+=== lora attempt 1 2026-08-14 01:22:08 ===
+base model_5000.pt | lora adapters 56 | trainable 1,598,400
+replay: 291 raw items from data/train_phase2b16k.bin (mixture ratio 0.50)
+train 582 eval 26 filtered 0

logs/sop_train.log

@@ -0,0 +1,42 @@
+train 407 eval 45
+loaded base model_final.pt (step 1043)
+step 25/600 loss 5.8748 282 tok/s
+  [eval 40] sft_val_loss 4.7163
+step 50/600 loss 2.4291 304 tok/s
+step 75/600 loss 3.0263 224 tok/s
+  [eval 80] sft_val_loss 4.5191
+step 100/600 loss 5.1817 225 tok/s
+  [eval 120] sft_val_loss 4.3853
+step 125/600 loss 2.2777 253 tok/s
+step 150/600 loss 4.0373 107 tok/s
+  [eval 160] sft_val_loss 4.2924
+step 175/600 loss 4.5904 272 tok/s
+step 200/600 loss 4.2869 282 tok/s
+  [eval 200] sft_val_loss 4.2100
+step 225/600 loss 2.6624 276 tok/s
+  [eval 240] sft_val_loss 4.1305
+step 250/600 loss 3.2132 270 tok/s
+step 275/600 loss 3.4927 271 tok/s
+  [eval 280] sft_val_loss 4.0843
+step 300/600 loss 3.7445 271 tok/s
+  [eval 320] sft_val_loss 4.0354
+step 325/600 loss 3.1865 250 tok/s
+step 350/600 loss 3.6414 149 tok/s
+  [eval 360] sft_val_loss 3.9930
+step 375/600 loss 4.2190 203 tok/s
+step 400/600 loss 2.5383 293 tok/s
+  [eval 400] sft_val_loss 3.9708
+step 425/600 loss 2.8048 289 tok/s
+  [eval 440] sft_val_loss 3.9236
+step 450/600 loss 3.2749 229 tok/s
+step 475/600 loss 3.3365 206 tok/s
+  [eval 480] sft_val_loss 3.9014
+step 500/600 loss 3.5851 256 tok/s
+  [eval 520] sft_val_loss 3.8652
+step 525/600 loss 2.5789 55 tok/s
+step 550/600 loss 3.3948 261 tok/s
+  [eval 560] sft_val_loss 3.8510
+step 575/600 loss 3.8072 235 tok/s
+step 600/600 loss 3.5173 291 tok/s
+  [eval 600] sft_val_loss 3.8381
+done -> ckpt/sop

logs/stage_a_orchestrator.log

@@ -0,0 +1 @@
+[stage_a] Wed Aug  5 21:19:40 CDT 2026 waiting for continue-pretrain (train_lm.py on tiny18m2) to exit...

logs/stage_a_sft_25m.log

@@ -0,0 +1,15 @@
+loaded base ckpt/tiny25m/model_5000.pt trainable 1723136/25426688 scope=last2
+train 154 eval 16 filtered 0
+step 25/114 loss 6.0848 sft 6.0576 181 tok/s
+  [eval 25] sft_val_loss 5.9019 val_ppl 13.22 [best ppl] [new best]
+  sample: Checklisting (s), 3-0xmodb, walking in a nearby nearby forest, and rangegegefa, hopped towards him, to the sky sky was coming from the air!
+step 50/114 loss 5.1076 sft 5.0472 166 tok/s
+  [eval 50] sft_val_loss 5.2695 val_ppl 8.31 [best ppl] [new best]
+  sample: Checklists) identified to a man who was in the (bd, 3-204]) +>>>[cpath, 55, *kwargr, 213, 1976
+step 75/114 loss 5.4804 sft 5.4118 152 tok/s
+  [eval 75] sft_val_loss 4.9834 val_ppl 8.41 [new best]
+  sample: Checklists(m) +=="Eende, a 3-2]\name %s, ****kwargr, in rangege([0x), 2) in a data[
+step 100/114 loss 4.7840 sft 4.7062 144 tok/s
+  [eval 100] sft_val_loss 4.8050 val_ppl 10.24 [new best]
+  sample: Checklists) identifiable entity or event; it requires a source beyond the claim itself. Verse refututututed from eback of the 209% change, and ran off to their in the
+done -> ckpt/tiny25m_sft_a best_sft=4.8050 best_ppl=8.31

logs/stage_v16k_runner.log

@@ -0,0 +1,38 @@
+== 1/2 vocab expansion: 8k -> 16k embeddings (exact-match mapping) ==
+base: ckpt/hybrid50m_pretrain/model_5000.pt  vocab 8192 -> 16384, 50.07M -> 52.70M
+token map: exact=8010 partial=182 noise=0 (97.8% exact)
+saved ckpt/hybrid50m_v16k_init.pt
+== 2/2 continue-pretrain on re-encoded phase-2b (16k) ==
+resumed ckpt/hybrid50m_v16k_pretrain/model_4000.pt at step 4000
+params: 52.92M  train_tokens: 31,238,190  steps_per_epoch: 30506
+step 4050/5000 loss 4.4874 lr 2.36e-05 119 tok/s
+step 4100/5000 loss 4.2918 lr 2.23e-05 145 tok/s
+step 4150/5000 loss 3.8822 lr 2.10e-05 119 tok/s
+step 4200/5000 loss 4.6073 lr 1.98e-05 121 tok/s
+step 4250/5000 loss 2.6758 lr 1.86e-05 119 tok/s
+step 4300/5000 loss 4.1535 lr 1.75e-05 120 tok/s
+step 4350/5000 loss 4.9097 lr 1.65e-05 119 tok/s
+step 4400/5000 loss 4.9595 lr 1.56e-05 130 tok/s
+step 4450/5000 loss 5.0700 lr 1.47e-05 142 tok/s
+step 4500/5000 loss 4.4885 lr 1.39e-05 165 tok/s
+  [eval step 4500] val_loss 3.0942 (best 3.0942)
+  [gen] Once upon a time,  there was a big house with many colors.
+  [gen] The little girl wanted to  ask for a moment and she did not see the way.
+  [gen] In the dark forest,  Tim was very happy that he could not find his new friend to get it again.
+  saved ckpt/hybrid50m_v16k_pretrain/model_4500.pt
+step 4550/5000 loss 3.7008 lr 1.31e-05 129 tok/s
+step 4600/5000 loss 4.1152 lr 1.25e-05 145 tok/s
+step 4650/5000 loss 4.2711 lr 1.19e-05 132 tok/s
+step 4700/5000 loss 4.1466 lr 1.14e-05 137 tok/s
+step 4750/5000 loss 4.1178 lr 1.10e-05 133 tok/s
+step 4800/5000 loss 3.8826 lr 1.06e-05 133 tok/s
+step 4850/5000 loss 4.1973 lr 1.04e-05 132 tok/s
+step 4900/5000 loss 4.4788 lr 1.02e-05 135 tok/s
+step 4950/5000 loss 4.8147 lr 1.00e-05 131 tok/s
+step 5000/5000 loss 3.8949 lr 1.00e-05 137 tok/s
+  [eval step 5000] val_loss 3.0865 (best 3.0865)
+  [gen] Once upon a time,  there was a little girl named Lily. They were curious and knew the cat would learn new things to make them happy.
+  [gen] The little girl wanted to  do something new. He did not know what was inside the bird helped him look up high in the yard.
+  [gen] In the dark forest,  the animals were happy to have fun with their friends. They were having fun and played together all day long.
+  saved ckpt/hybrid50m_v16k_pretrain/model_5000.pt
+done

logs/tokenizer16k.log

@@ -0,0 +1,42 @@
+== 1/3 retrain 16k tokenizer on 528M-token corpus (streaming) ==
+
+
+
+saved data/tokenizer16k.json  vocab=16384  lines=11,545,267  630s
+== 2/3 re-encode phase-2b corpus (32.5M) ==
+data/train_phase2b16k.bin: 32,562,432 old tokens -> 31,238,190 new tokens (0.96x) in 110s
+== 3/3 re-encode full corpus (528M) ==
+== 3/3 (resumed) re-encode full corpus (528M) ==
+== 3/3 (2nd resume) re-encode full corpus (528M, chunked fix) ==
+=== full re-encode attempt 1 2026-08-11 12:24:59 ===
+=== full re-encode attempt 1 2026-08-11 23:21:56 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1364s
+=== attempt 1 exit 0 23:44:42 ===
+attempt 1 lines=11544766 expect=11545267
+incomplete (or killed); retrying in 10s
+=== full re-encode attempt 2 2026-08-11 23:44:56 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1241s
+=== attempt 2 exit 0 00:05:38 ===
+attempt 2 lines=11544766 expect=11545267
+incomplete (or killed); retrying in 10s
+=== full re-encode attempt 3 2026-08-12 00:05:51 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1189s
+=== attempt 3 exit 0 00:25:41 ===
+attempt 3 lines=11544766 expect=11545267
+incomplete (or killed); retrying in 10s
+=== full re-encode attempt 4 2026-08-12 00:25:54 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1174s
+=== attempt 4 exit 0 00:45:29 ===
+attempt 4 lines=11544766 expect=11545267
+incomplete (or killed); retrying in 10s
+=== full re-encode attempt 5 2026-08-12 00:45:42 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1187s
+=== attempt 5 exit 0 01:05:31 ===
+attempt 5 lines=11544766 expect=11545267
+incomplete (or killed); retrying in 10s
+=== full re-encode attempt 6 2026-08-12 01:05:44 ===
+=== full re-encode attempt 1 2026-08-12 01:13:44 ===
+data/train_full16k.bin: 528,000,852 old tokens -> 520,133,183 new tokens (0.99x) in 1503s
+=== attempt 1 exit 0 01:38:49 ===
+attempt 1 lines=11544766 expect=11544766 tokens=520133183 expect_tokens=520133183
+full re-encode COMPLETE (11544766 lines, 520133183 tokens)

logs/train_domain.log

@@ -0,0 +1,108 @@
+init-from ckpt/nlp_full/model_9000.pt (weights only, fresh schedule)
+params: 7.79M  train_tokens: 1,935,861  steps_per_epoch: 472
+step 1/1600 loss 7.6557 lr 1.00e-04 26168 tok/s
+step 25/1600 loss 5.3070 lr 1.00e-04 1010 tok/s
+step 50/1600 loss 4.1954 lr 1.00e-04 943 tok/s
+step 75/1600 loss 3.5035 lr 9.99e-05 946 tok/s
+step 100/1600 loss 3.0055 lr 9.98e-05 926 tok/s
+step 125/1600 loss 3.0248 lr 9.95e-05 920 tok/s
+step 150/1600 loss 2.3517 lr 9.91e-05 916 tok/s
+step 175/1600 loss 2.6255 lr 9.86e-05 907 tok/s
+step 200/1600 loss 1.9153 lr 9.79e-05 907 tok/s
+  [eval step 200] val_loss 2.2379 (best 2.2379)
+  [gen] Once upon a time,  there lived a little fish. They were sad because they loved to play with his friends.
+  [gen] The little girl wanted to  help. The girl was lost her toy dog with his paws. Tim had an idea. She said, she made a big hug and how. He gave them all their toys. But the boy was not happy. He was sad.
+  [gen] In the dark forest,  they were playing, and they could not find his friends to play, they played together and had a fun day on on. They lived in a small house, and they would take care of their friends and played with his toy car, and Tim and Max loved to play, Spot was under trees, and the kids in the place he always made people happy when it was.
+  saved ckpt/nlp_domain/model_200.pt
+step 225/1600 loss 2.2984 lr 9.72e-05 574 tok/s
+step 250/1600 loss 2.0293 lr 9.64e-05 916 tok/s
+step 275/1600 loss 1.8629 lr 9.54e-05 918 tok/s
+step 300/1600 loss 1.9641 lr 9.43e-05 914 tok/s
+step 325/1600 loss 2.3289 lr 9.32e-05 916 tok/s
+step 350/1600 loss 2.9401 lr 9.19e-05 869 tok/s
+step 375/1600 loss 1.4854 lr 9.06e-05 911 tok/s
+step 400/1600 loss 2.5227 lr 8.91e-05 961 tok/s
+  [eval step 400] val_loss 1.9466 (best 1.9466)
+  [gen] Once upon a time,  in the house, there was a red apple on her toy car on his mom. She would put it at it under her box with her toy car and she went to it, Sue picked up the cake, and had so much fun. She could not get the toy box, she got on the table with the big toy toy box, but he needed it very much. He wanted to play with
+  [gen] The little girl wanted to  find the cat. The bird and a rabbit. The duck was in the tree.
+  [gen] In the dark forest,  they saw a little bird. The bird was scared to know what was looking for the bird. He was not happy. But when he could take the ball, it did not get any water at all. But it had an idea. He was excited and happy to find his friend. His friends in the park with his friends. They were tired from the park to play with the ball and make them
+  saved ckpt/nlp_domain/model_400.pt
+step 425/1600 loss 2.0663 lr 8.76e-05 596 tok/s
+step 450/1600 loss 2.0553 lr 8.60e-05 959 tok/s
+step 475/1600 loss 1.7122 lr 8.43e-05 975 tok/s
+step 500/1600 loss 1.9163 lr 8.25e-05 979 tok/s
+step 525/1600 loss 2.0934 lr 8.07e-05 976 tok/s
+step 550/1600 loss 1.4521 lr 7.88e-05 963 tok/s
+step 575/1600 loss 2.1106 lr 7.68e-05 963 tok/s
+step 600/1600 loss 2.2216 lr 7.48e-05 965 tok/s
+  [eval step 600] val_loss 1.8222 (best 1.8222)
+  [gen] Once upon a time,  interhh, "Sast of a real-sar"'s to it has to be no one.
+  [gen] The little girl wanted to  have a letter to the bird. The bird saw the tree and told the bird, so he took the bird to Tim's dog with his friends.
+  [gen] In the dark forest,  the bird came into the sun. The bird saw a big tree in the middle of the water. He had many toys inside the park and it was very fast to play with it. They decided to be there a dog with his friend, the cat had a ball of the dog. They went into the tree and they sat down together on the ground it.
+  saved ckpt/nlp_domain/model_600.pt
+step 625/1600 loss 1.9078 lr 7.27e-05 582 tok/s
+step 650/1600 loss 2.2048 lr 7.06e-05 893 tok/s
+step 675/1600 loss 2.6666 lr 6.85e-05 884 tok/s
+step 700/1600 loss 1.5878 lr 6.63e-05 991 tok/s
+step 725/1600 loss 1.8886 lr 6.41e-05 1034 tok/s
+step 750/1600 loss 1.3845 lr 6.18e-05 1035 tok/s
+step 775/1600 loss 1.6045 lr 5.96e-05 1033 tok/s
+step 800/1600 loss 1.8970 lr 5.73e-05 1059 tok/s
+  [eval step 800] val_loss 1.7502 (best 1.7502)
+  [gen] Once upon a time,  in the park, they all laughed and playing outside, and they were very happy to have fun with their friends in his friends, and they lived happily together every day long.
+  [gen] The little girl wanted to  play with the bird. The cat said, "I am here?" His friends had an in the tree. It was a lot of log. The bird came up and jumped out from behind the low it. As he got closer, there were many more people all the small friends. One day, a little boy named for the little girl said to the bird, "Tim that the bird could fly
+  [gen] In the dark forest,  the frog came to and saw the tree, in the sun, in the green sky on the trees, he saw a big box on it. The rock was not all the new, it made it up with his box. He looked at the door and put it around the room in the hole in the tree, but his friends had red came to the woods to look for him to get the house
+  saved ckpt/nlp_domain/model_800.pt
+step 825/1600 loss 1.9430 lr 5.50e-05 601 tok/s
+step 850/1600 loss 1.9832 lr 5.27e-05 1023 tok/s
+step 875/1600 loss 2.3119 lr 5.04e-05 1026 tok/s
+step 900/1600 loss 2.0030 lr 4.82e-05 1012 tok/s
+step 925/1600 loss 1.7119 lr 4.59e-05 1042 tok/s
+step 950/1600 loss 1.6022 lr 4.37e-05 1017 tok/s
+step 975/1600 loss 1.9157 lr 4.15e-05 1011 tok/s
+step 1000/1600 loss 1.6265 lr 3.94e-05 1006 tok/s
+  [eval step 1000] val_loss 1.7034 (best 1.7034)
+  [gen] Once upon a time,  in a small house with lots of toys from the sun. Tim loved to play in the park. One day, Tom played on the yard with his friends. They had fun and were very nice to play with them.
+  [gen] The little girl wanted to  help the bird. The cat liked the tree. The little blue frog lived near a big tree. The small box on his yard. One day, the fox saw a sad dog. The bird had a lot of them. The toy car was a shiny ball. He did not know what it was. Tim wanted to play with the mouse.
+  [gen] In the dark forest,  there was a big tree in the low air behind the sky. He wanted to play near it, he saw an idea. Tim thought his toy box to Sam and said, "Wow! I can't find it. They got up and opened it and they found it down and started to reach the pond. They went back to the ball, and they all had lots of friends playing in the park
+  saved ckpt/nlp_domain/model_1000.pt
+step 1025/1600 loss 1.8698 lr 3.73e-05 625 tok/s
+step 1050/1600 loss 1.7240 lr 3.52e-05 989 tok/s
+step 1075/1600 loss 1.8379 lr 3.32e-05 1005 tok/s
+step 1100/1600 loss 1.3732 lr 3.12e-05 1011 tok/s
+step 1125/1600 loss 1.6244 lr 2.93e-05 1008 tok/s
+step 1150/1600 loss 1.6726 lr 2.75e-05 1002 tok/s
+step 1175/1600 loss 1.6783 lr 2.57e-05 966 tok/s
+step 1200/1600 loss 1.6297 lr 2.40e-05 948 tok/s
+  [eval step 1200] val_loss 1.6723 (best 1.6723)
+  [gen] Once upon a time,  there was a little girl named Lily. Sue loved to play by the woods on her room because she had many pretty colors in his toys with him. One day, he had a plan to be in his home with his friends.
+  [gen] The little girl wanted to  play outside. The bird was happy and couldn't wait to have all the friends he were not seen anymore because of his big dog. The cat would not go as Tims with his red ball, but he had no back to place to play with it with one day, they knew how to be in his house near him.
+  [gen] In the dark forest,  the cat and the bird was a boy went to into the house. The cat saw his friend, the dog, and asked Tim that his mom asked Max for a walk in the park with his friends in the lake with his home when he could not find the little bird he had it on her friends to play outside.
+  saved ckpt/nlp_domain/model_1200.pt
+step 1225/1600 loss 1.9096 lr 2.24e-05 624 tok/s
+step 1250/1600 loss 1.9304 lr 2.09e-05 1044 tok/s
+step 1275/1600 loss 2.3165 lr 1.94e-05 1042 tok/s
+step 1300/1600 loss 1.3123 lr 1.81e-05 1043 tok/s
+step 1325/1600 loss 2.0325 lr 1.68e-05 1045 tok/s
+step 1350/1600 loss 1.6920 lr 1.57e-05 1044 tok/s
+step 1375/1600 loss 1.4900 lr 1.46e-05 1022 tok/s
+step 1400/1600 loss 1.5946 lr 1.36e-05 1036 tok/s
+  [eval step 1400] val_loss 1.6536 (best 1.6536)
+  [gen] Once upon a time,  in a fun at the green grass. They all had red ball of toys with their family every day on night long. They liked to play together all day long. One day, they played with her friends all day long, making everyone happy and they played with his friends. And Tim became best friends always listened all day long in their tree friends.
+  [gen] The little girl wanted to  help the little girl.
+  [gen] In the dark forest,  they saw a big box in the living room with it and made him open it. Tim was very happy to see his new friends playing with their toy dog on it. The bird had many friends to play with the tree all day long. They played together every day, and they were not so much fun outside their own friends to help them. And they lived happily ever after it was a time in
+  saved ckpt/nlp_domain/model_1400.pt
+step 1425/1600 loss 1.4715 lr 1.28e-05 628 tok/s
+step 1450/1600 loss 1.8082 lr 1.21e-05 1035 tok/s
+step 1475/1600 loss 1.6201 lr 1.14e-05 1021 tok/s
+step 1500/1600 loss 1.3674 lr 1.09e-05 1030 tok/s
+step 1525/1600 loss 1.6458 lr 1.05e-05 1032 tok/s
+step 1550/1600 loss 1.6130 lr 1.02e-05 1038 tok/s
+step 1575/1600 loss 1.8113 lr 1.01e-05 1038 tok/s
+step 1600/1600 loss 1.1391 lr 1.00e-05 1038 tok/s
+  [eval step 1600] val_loss 1.6435 (best 1.6435)
+  [gen] Once upon a time,  there was a little dog. They learned to play on the big fish and the cat. They would always told the bird how to rise, they were all the friends in the forest, and they all had many days in the fun fun day.
+  [gen] The little girl wanted to  eat the bird. The dog saw a cat in the tree. The flower was in the tree's tree. It had many people in the tree! All the fish went back home, having so much fun day for it everywhere. Every story how to rise, the bear had a big bowl of his mouth with many other birds, and trees in the sun all around the box when he wanted to eat
+  [gen] In the dark forest,  there was a big storm came to the field with them with him in the sun. Tim loved to play and look for friends for his friend to help. One day, Max saw that every other people are friends who had to give his new friends to her friends who had many better friends. They played together all day long, and Spot liked to play all day on, and they would go all their
+  saved ckpt/nlp_domain/model_1600.pt
+done

logs/train_dpo_lora.log

@@ -0,0 +1,21 @@
+adapter-only DPO: 56 lora adapters, trainable 1,156,032
+loaded 3004 preference pairs; base ckpt/tiny25m/model_best.pt
+step 10/751 dpo_loss 0.6726
+step 20/751 dpo_loss 0.6363
+  [eval 25] val_ppl 8.68 guard 60.0 [best ppl]
+step 30/751 dpo_loss 0.6138
+step 40/751 dpo_loss 0.5883
+step 50/751 dpo_loss 0.5716
+  [eval 50] val_ppl 7.85 guard 60.0 [best ppl]
+step 60/751 dpo_loss 0.5132
+step 70/751 dpo_loss 0.4127
+  [eval 75] val_ppl 5.82 guard 60.0 [best ppl]
+step 80/751 dpo_loss 0.3410
+step 90/751 dpo_loss 0.2706
+step 100/751 dpo_loss 0.0957
+  [eval 100] val_ppl 9.46 guard 60.0
+  saved model_100.pt
+step 110/751 dpo_loss 0.0862
+step 120/751 dpo_loss 0.0791
+  [eval 125] val_ppl 72.71 guard 60.0
+PPL GUARD BREACH — aborting (last good state preserved)

logs/train_full.log

@@ -0,0 +1,301 @@
+resumed ckpt/nlp_full/model_3250.pt at step 3250
+params: 7.79M  train_tokens: 528,000,852  steps_per_epoch: 128906
+step 3275/8250 loss 3.2652 lr 1.05e-04 227 tok/s
+step 3300/8250 loss 3.0516 lr 1.05e-04 64 tok/s
+step 3325/8250 loss 3.0503 lr 1.04e-04 34 tok/s
+step 3350/8250 loss 3.1550 lr 1.03e-04 272 tok/s
+step 3375/8250 loss 3.1012 lr 1.03e-04 147 tok/s
+step 3400/8250 loss 3.1201 lr 1.02e-04 144 tok/s
+step 3425/8250 loss 3.0962 lr 1.02e-04 148 tok/s
+step 3450/8250 loss 3.0490 lr 1.01e-04 151 tok/s
+step 3475/8250 loss 3.2165 lr 1.00e-04 184 tok/s
+step 3500/8250 loss 2.8395 lr 9.95e-05 231 tok/s
+  [eval step 3500] val_loss 3.0141 (best 3.0141)
+  [gen] Once upon a time,  there lived a best friend. From that day on, Lucy met a special surprise.
+  [gen] The little girl wanted to  eat the fruit. And it felt guilty for the boat.
+  [gen] In the dark forest,  there was a little boy who liked to play in an unusual place.
+  saved ckpt/nlp_full/model_3500.pt
+step 3525/8250 loss 2.8705 lr 9.89e-05 487 tok/s
+step 3550/8250 loss 3.0629 lr 9.82e-05 738 tok/s
+step 3575/8250 loss 3.0515 lr 9.76e-05 779 tok/s
+step 3600/8250 loss 3.1774 lr 9.69e-05 252 tok/s
+step 3625/8250 loss 2.9818 lr 9.62e-05 525 tok/s
+step 3650/8250 loss 2.9059 lr 9.56e-05 30 tok/s
+step 3675/8250 loss 3.0357 lr 9.49e-05 547 tok/s
+step 3700/8250 loss 2.9865 lr 9.42e-05 87 tok/s
+step 3725/8250 loss 2.9513 lr 9.36e-05 133 tok/s
+step 3750/8250 loss 2.8726 lr 9.29e-05 702 tok/s
+step 3775/8250 loss 2.9917 lr 9.22e-05 730 tok/s
+step 3800/8250 loss 3.1404 lr 9.16e-05 732 tok/s
+step 3825/8250 loss 2.7713 lr 9.09e-05 734 tok/s
+step 3850/8250 loss 2.9750 lr 9.02e-05 735 tok/s
+step 3875/8250 loss 2.8550 lr 8.95e-05 729 tok/s
+step 3900/8250 loss 2.8402 lr 8.89e-05 483 tok/s
+step 3925/8250 loss 2.7494 lr 8.82e-05 727 tok/s
+step 3950/8250 loss 2.9591 lr 8.75e-05 680 tok/s
+step 3975/8250 loss 3.1924 lr 8.68e-05 170 tok/s
+step 4000/8250 loss 2.9679 lr 8.61e-05 138 tok/s
+  [eval step 4000] val_loss 2.8675 (best 2.8675)
+  [gen] Once upon a time,  there was a happy place in the room. One day, Bob had an idea. They played with his toy car and had lots of fun. As they were playing together, they played all day long time. Tim loved to play with his toys. They loved to have fun more friends forever.
+  [gen] The little girl wanted to  play. The dog wants to help the bird. He is happy that he was not the old man. They look, but they were very fast. The dog was sad.
+  [gen] In the dark forest,  there were many trees and trees. They liked to play together every day on his big tree. They became very good friends.
+  saved ckpt/nlp_full/model_4000.pt
+step 4025/8250 loss 2.9230 lr 8.55e-05 213 tok/s
+step 4050/8250 loss 2.9135 lr 8.48e-05 556 tok/s
+resumed ckpt/nlp_full/model_4000.pt at step 4000
+params: 7.79M  train_tokens: 528,000,852  steps_per_epoch: 128906
+step 4025/9000 loss nan lr 9.43e-05 823 tok/s
+step 4050/9000 loss nan lr 9.37e-05 822 tok/s
+resumed ckpt/nlp_full/model_4000.pt at step 4000
+params: 7.79M  train_tokens: 528,000,852  steps_per_epoch: 128906
+step 4025/9000 loss 2.9245 lr 9.43e-05 903 tok/s
+step 4050/9000 loss 2.9155 lr 9.37e-05 914 tok/s
+step 4075/9000 loss 2.9106 lr 9.30e-05 912 tok/s
+step 4100/9000 loss 2.9309 lr 9.24e-05 911 tok/s
+step 4125/9000 loss 2.9195 lr 9.18e-05 913 tok/s
+step 4150/9000 loss 2.7386 lr 9.12e-05 909 tok/s
+step 4175/9000 loss 2.9189 lr 9.06e-05 915 tok/s
+step 4200/9000 loss 2.8414 lr 9.00e-05 911 tok/s
+step 4225/9000 loss 2.7259 lr 8.93e-05 916 tok/s
+step 4250/9000 loss 2.9817 lr 8.87e-05 888 tok/s
+step 4275/9000 loss 2.8962 lr 8.81e-05 231 tok/s
+step 4300/9000 loss 2.9390 lr 8.75e-05 104 tok/s
+step 4325/9000 loss 2.9470 lr 8.69e-05 962 tok/s
+step 4350/9000 loss 2.6792 lr 8.62e-05 474 tok/s
+step 4375/9000 loss 2.8925 lr 8.56e-05 434 tok/s
+step 4400/9000 loss 2.9679 lr 8.50e-05 953 tok/s
+step 4425/9000 loss 3.0739 lr 8.44e-05 960 tok/s
+step 4450/9000 loss 2.7087 lr 8.37e-05 931 tok/s
+step 4475/9000 loss 2.9035 lr 8.31e-05 933 tok/s
+step 4500/9000 loss 2.7015 lr 8.25e-05 932 tok/s
+  [eval step 4500] val_loss 2.7695 (best 2.7695)
+  [gen] Once upon a time,  in a big park, and had lots of fun. One day, Lily saw a big puddle on the wall full of juice! Sue was very happy when she went outside to play with it with her toy cars and it would make yummy food for them.
+  [gen] The little girl wanted to  help the girl. The boy and the dog. He was sad.
+  [gen] In the dark forest,  there was a big, green leaves. As they were walking around to visit the other kids and said, "Yes!" They put on the paper and laughed at their mom. They felt happy.
+  saved ckpt/nlp_full/model_4500.pt
+step 4525/9000 loss 2.8634 lr 8.19e-05 583 tok/s
+step 4550/9000 loss 2.7695 lr 8.12e-05 934 tok/s
+step 4575/9000 loss 2.8348 lr 8.06e-05 931 tok/s
+step 4600/9000 loss 2.6004 lr 8.00e-05 929 tok/s
+step 4625/9000 loss 2.8610 lr 7.94e-05 924 tok/s
+step 4650/9000 loss 2.9892 lr 7.88e-05 918 tok/s
+step 4675/9000 loss 2.8633 lr 7.81e-05 930 tok/s
+step 4700/9000 loss 2.8188 lr 7.75e-05 929 tok/s
+step 4725/9000 loss 2.9134 lr 7.69e-05 929 tok/s
+step 4750/9000 loss 2.7791 lr 7.63e-05 925 tok/s
+step 4775/9000 loss 2.8376 lr 7.56e-05 925 tok/s
+step 4800/9000 loss 2.5659 lr 7.50e-05 925 tok/s
+step 4825/9000 loss 2.9143 lr 7.44e-05 911 tok/s
+step 4850/9000 loss 2.6105 lr 7.38e-05 905 tok/s
+step 4875/9000 loss 2.7211 lr 7.31e-05 97 tok/s
+step 4900/9000 loss 2.8035 lr 7.25e-05 262 tok/s
+step 4925/9000 loss 2.9030 lr 7.19e-05 878 tok/s
+step 4950/9000 loss 2.5854 lr 7.13e-05 929 tok/s
+step 4975/9000 loss 2.8602 lr 7.07e-05 1023 tok/s
+step 5000/9000 loss 2.7809 lr 7.00e-05 996 tok/s
+  [eval step 5000] val_loss 2.7009 (best 2.7009)
+  [gen] Once upon a time,  in a big house, there was a little box in many colorful things. One day, Bob had a big, shiny hat that he always wanted to share the toys and eat yummy apple.
+  [gen] The little girl wanted to  see the boy. The girl had an idea and asked his friend, a little girl who came to help. She was sad because she did not want to be brave. He wanted to do it and didn't know what to do.
+  [gen] In the dark forest,  there was a very impressive rock. One day, while playing with his toy box in a pretty house. He thought it would make a mess with a big ball.
+  saved ckpt/nlp_full/model_5000.pt
+step 5025/9000 loss 2.5899 lr 6.94e-05 620 tok/s
+step 5050/9000 loss 2.6278 lr 6.88e-05 986 tok/s
+step 5075/9000 loss 2.7113 lr 6.82e-05 859 tok/s
+step 5100/9000 loss 2.6565 lr 6.76e-05 872 tok/s
+step 5125/9000 loss 2.6728 lr 6.70e-05 960 tok/s
+step 5150/9000 loss 2.7385 lr 6.63e-05 971 tok/s
+step 5175/9000 loss 2.7077 lr 6.57e-05 948 tok/s
+step 5200/9000 loss 2.7743 lr 6.51e-05 910 tok/s
+step 5225/9000 loss 2.7293 lr 6.45e-05 911 tok/s
+step 5250/9000 loss 2.6442 lr 6.39e-05 915 tok/s
+step 5275/9000 loss 2.4881 lr 6.33e-05 909 tok/s
+step 5300/9000 loss 2.6077 lr 6.27e-05 895 tok/s
+step 5325/9000 loss 2.6470 lr 6.21e-05 893 tok/s
+step 5350/9000 loss 2.5361 lr 6.15e-05 858 tok/s
+step 5375/9000 loss 2.5995 lr 6.09e-05 897 tok/s
+step 5400/9000 loss 2.5962 lr 6.03e-05 899 tok/s
+step 5425/9000 loss 3.0198 lr 5.97e-05 880 tok/s
+step 5450/9000 loss 2.8351 lr 5.91e-05 904 tok/s
+step 5475/9000 loss 2.6752 lr 5.85e-05 890 tok/s
+step 5500/9000 loss 2.6133 lr 5.79e-05 895 tok/s
+  [eval step 5500] val_loss 2.6462 (best 2.6462)
+  [gen] Once upon a time,  in a small room, there lived near the park. One day, Fluffy flew far away, and Sam found a big red ball in the window.
+  [gen] The little girl wanted to  see the little girl, so she asked her mom, "Why are you too strong and I will not have any way to eat it."
+  [gen] In the dark forest,  there was a fun game in a big pond. The sun in the town loved to play and have catch it every day.
+  saved ckpt/nlp_full/model_5500.pt
+step 5525/9000 loss 2.7746 lr 5.73e-05 548 tok/s
+step 5550/9000 loss 2.3815 lr 5.67e-05 857 tok/s
+step 5575/9000 loss 2.7155 lr 5.61e-05 857 tok/s
+step 5600/9000 loss 2.5269 lr 5.55e-05 860 tok/s
+step 5625/9000 loss 2.7181 lr 5.50e-05 859 tok/s
+step 5650/9000 loss 2.5197 lr 5.44e-05 879 tok/s
+step 5675/9000 loss 2.5076 lr 5.38e-05 1021 tok/s
+step 5700/9000 loss 2.6206 lr 5.32e-05 990 tok/s
+step 5725/9000 loss 2.4553 lr 5.26e-05 997 tok/s
+step 5750/9000 loss 2.6641 lr 5.21e-05 1075 tok/s
+step 5775/9000 loss 2.6422 lr 5.15e-05 1121 tok/s
+step 5800/9000 loss 2.5006 lr 5.09e-05 1098 tok/s
+step 5825/9000 loss 2.6822 lr 5.04e-05 1118 tok/s
+step 5850/9000 loss 2.6610 lr 4.98e-05 1130 tok/s
+step 5875/9000 loss 2.5627 lr 4.92e-05 1150 tok/s
+step 5900/9000 loss 2.3672 lr 4.87e-05 944 tok/s
+step 5925/9000 loss 2.6929 lr 4.81e-05 920 tok/s
+step 5950/9000 loss 2.4623 lr 4.76e-05 985 tok/s
+step 5975/9000 loss 2.5862 lr 4.70e-05 953 tok/s
+step 6000/9000 loss 2.6821 lr 4.65e-05 975 tok/s
+  [eval step 6000] val_loss 2.6086 (best 2.6086)
+  [gen] Once upon a time,  in a small house and a big house with many toys, and flowers. One day, Lily found a toy car on it! Sue was very happy to play with their toys with Tim's toys every day.
+  [gen] The little girl wanted to  help the boy. But the bear was scared, but he could not find it. The boy saw the man and asked, "Why are you sad?"
+  [gen] In the dark forest,  there lived a nice lady named Kitty. Lily loved to explore together. One day, Lucy met a very little girl named Sue. Mia had a big toy dog named Max. Sam loved the cat, Max and Tim liked to play games in the park.
+  saved ckpt/nlp_full/model_6000.pt
+step 6025/9000 loss 2.6786 lr 4.59e-05 574 tok/s
+step 6050/9000 loss 2.5485 lr 4.54e-05 1036 tok/s
+step 6075/9000 loss 2.6821 lr 4.48e-05 1034 tok/s
+step 6100/9000 loss 2.6589 lr 4.43e-05 1017 tok/s
+step 6125/9000 loss 2.4764 lr 4.37e-05 1033 tok/s
+step 6150/9000 loss 2.6516 lr 4.32e-05 1039 tok/s
+step 6175/9000 loss 2.7046 lr 4.27e-05 1036 tok/s
+step 6200/9000 loss 2.7285 lr 4.22e-05 1015 tok/s
+step 6225/9000 loss 2.4805 lr 4.16e-05 1040 tok/s
+step 6250/9000 loss 2.3852 lr 4.11e-05 1092 tok/s
+step 6275/9000 loss 2.3647 lr 4.06e-05 1037 tok/s
+step 6300/9000 loss 2.7131 lr 4.01e-05 1045 tok/s
+step 6325/9000 loss 2.6772 lr 3.96e-05 1131 tok/s
+step 6350/9000 loss 2.5595 lr 3.91e-05 657 tok/s
+step 6375/9000 loss 2.4779 lr 3.86e-05 834 tok/s
+step 6400/9000 loss 2.5557 lr 3.81e-05 851 tok/s
+step 6425/9000 loss 2.6304 lr 3.76e-05 851 tok/s
+step 6450/9000 loss 2.7155 lr 3.71e-05 845 tok/s
+step 6475/9000 loss 2.7958 lr 3.66e-05 845 tok/s
+step 6500/9000 loss 2.4566 lr 3.61e-05 853 tok/s
+  [eval step 6500] val_loss 2.5781 (best 2.5781)
+  [gen] Once upon a time,  in a big pond near her house. She loved to slide down the grass, and it was very sweet, but one day on, she would bounce on it, and forth, and they were having fun together.
+  [gen] The little girl wanted to  help the big boy. The cat asked the little bird, "Can I use my nest to play with you?" The bird wanted to help, but he remembered that the ball was not happy. It could not listen to his friends.
+  [gen] In the dark forest,  there was a big, yummy cake. They all loved to play and have fun in the park.
+  saved ckpt/nlp_full/model_6500.pt
+step 6525/9000 loss 2.5647 lr 3.56e-05 544 tok/s
+step 6550/9000 loss 2.4096 lr 3.51e-05 843 tok/s
+step 6575/9000 loss 2.7280 lr 3.46e-05 846 tok/s
+step 6600/9000 loss 2.5285 lr 3.42e-05 849 tok/s
+step 6625/9000 loss 2.4943 lr 3.37e-05 864 tok/s
+step 6650/9000 loss 2.6783 lr 3.32e-05 861 tok/s
+step 6675/9000 loss 2.5472 lr 3.28e-05 862 tok/s
+step 6700/9000 loss 2.8120 lr 3.23e-05 854 tok/s
+step 6725/9000 loss 2.5693 lr 3.18e-05 857 tok/s
+step 6750/9000 loss 2.5588 lr 3.14e-05 860 tok/s
+step 6775/9000 loss 2.5237 lr 3.09e-05 852 tok/s
+step 6800/9000 loss 2.5752 lr 3.05e-05 840 tok/s
+step 6825/9000 loss 2.7470 lr 3.01e-05 839 tok/s
+step 6850/9000 loss 2.4375 lr 2.96e-05 844 tok/s
+step 6875/9000 loss 2.5079 lr 2.92e-05 838 tok/s
+step 6900/9000 loss 2.6921 lr 2.88e-05 838 tok/s
+step 6925/9000 loss 2.5621 lr 2.83e-05 841 tok/s
+step 6950/9000 loss 2.5515 lr 2.79e-05 833 tok/s
+step 6975/9000 loss 2.6248 lr 2.75e-05 839 tok/s
+step 7000/9000 loss 2.5410 lr 2.71e-05 839 tok/s
+  [eval step 7000] val_loss 2.5547 (best 2.5547)
+  [gen] Once upon a time,  there was a big, red ball that he loved to play outside in the yard and eat it every day long. He would jump and look all around the trees, around through the deep green woods, and they were playing in the sun and having so much fun.
+  [gen] The little girl wanted to  help the bird, but the cat was sad. The dog said, "Hi, bird! I made a new friend. Can you show me my ball back to your nest."
+  [gen] In the dark forest,  in the forest was full of things to see who lived near. The bird had many friends for food and friends. They liked to play together every day in the garden all day long and were very excited to play with them.
+  saved ckpt/nlp_full/model_7000.pt
+step 7025/9000 loss 2.5539 lr 2.67e-05 539 tok/s
+step 7050/9000 loss 2.5151 lr 2.63e-05 826 tok/s
+step 7075/9000 loss 2.6173 lr 2.59e-05 833 tok/s
+step 7100/9000 loss 2.5911 lr 2.55e-05 826 tok/s
+step 7125/9000 loss 2.7414 lr 2.51e-05 822 tok/s
+step 7150/9000 loss 2.6786 lr 2.47e-05 819 tok/s
+step 7175/9000 loss 2.6262 lr 2.43e-05 815 tok/s
+step 7200/9000 loss 2.4232 lr 2.40e-05 171 tok/s
+step 7225/9000 loss 2.6298 lr 2.36e-05 124 tok/s
+step 7250/9000 loss 2.3896 lr 2.32e-05 36 tok/s
+step 7275/9000 loss 2.6787 lr 2.29e-05 32 tok/s
+step 7300/9000 loss 2.7686 lr 2.25e-05 21 tok/s
+step 7325/9000 loss 2.5575 lr 2.21e-05 41 tok/s
+step 7350/9000 loss 2.4127 lr 2.18e-05 168 tok/s
+step 7375/9000 loss 2.5340 lr 2.15e-05 63 tok/s
+step 7400/9000 loss 2.4636 lr 2.11e-05 271 tok/s
+step 7425/9000 loss 2.5265 lr 2.08e-05 960 tok/s
+step 7450/9000 loss 2.5189 lr 2.04e-05 1031 tok/s
+step 7475/9000 loss 2.6430 lr 2.01e-05 1027 tok/s
+step 7500/9000 loss 2.4478 lr 1.98e-05 1018 tok/s
+  [eval step 7500] val_loss 2.5380 (best 2.5380)
+  [gen] Once upon a time,  in a small house with a big yard. One day, and the rock would hop all by a big tree every day.
+  [gen] The little girl wanted to  help the fish, and they both became friends. They played together every day, and And they lived happily ever after, and they all played together with their new friends.
+  [gen] In the dark forest,  the people saw a soft hole near the hill. The big tree was very happy to watch it play with the ball all day long. He would fly high into the air and look for toys in the water. His friends loved his toy car very fast, so they played with the ball in the sand all day long, and Tim made friends and played with the red ball every day. And from that
+  saved ckpt/nlp_full/model_7500.pt
+step 7525/9000 loss 2.3349 lr 1.95e-05 601 tok/s
+step 7550/9000 loss 2.5411 lr 1.92e-05 977 tok/s
+step 7575/9000 loss 2.7956 lr 1.89e-05 956 tok/s
+step 7600/9000 loss 2.4876 lr 1.86e-05 960 tok/s
+step 7625/9000 loss 2.4813 lr 1.83e-05 940 tok/s
+step 7650/9000 loss 2.6539 lr 1.80e-05 952 tok/s
+step 7675/9000 loss 2.4810 lr 1.77e-05 961 tok/s
+step 7700/9000 loss 2.2995 lr 1.74e-05 944 tok/s
+step 7725/9000 loss 2.4481 lr 1.71e-05 917 tok/s
+step 7750/9000 loss 2.5524 lr 1.69e-05 894 tok/s
+step 7775/9000 loss 2.5015 lr 1.66e-05 1027 tok/s
+step 7800/9000 loss 2.7047 lr 1.63e-05 1020 tok/s
+step 7825/9000 loss 2.4839 lr 1.61e-05 1029 tok/s
+step 7850/9000 loss 2.4050 lr 1.58e-05 1029 tok/s
+step 7875/9000 loss 2.5614 lr 1.56e-05 1033 tok/s
+step 7900/9000 loss 2.6719 lr 1.53e-05 1030 tok/s
+step 7925/9000 loss 2.7266 lr 1.51e-05 1023 tok/s
+step 7950/9000 loss 2.4625 lr 1.49e-05 1028 tok/s
+step 7975/9000 loss 2.6055 lr 1.46e-05 1007 tok/s
+step 8000/9000 loss 2.5579 lr 1.44e-05 1011 tok/s
+  [eval step 8000] val_loss 2.5261 (best 2.5261)
+  [gen] Once upon a time,  in a big dark forest, there lived a blue box in the sky. The red ball liked to play by the big red car on it every day long. One sunny day, Tom and Max played together all day long.
+  [gen] The little girl wanted to  help the little girl. The boy had a big, fluffy cat with a small toy all day.
+  [gen] In the dark forest,  there lived a little girl named Bob. The cat lived in a big tree with many friends. The blue house had a friend called to play and have fun together. The big red ball would fly high in one day, playing near the water that could grow.
+  saved ckpt/nlp_full/model_8000.pt
+step 8025/9000 loss 2.5474 lr 1.42e-05 627 tok/s
+step 8050/9000 loss 2.5776 lr 1.40e-05 982 tok/s
+step 8075/9000 loss 2.5785 lr 1.38e-05 996 tok/s
+step 8100/9000 loss 2.5373 lr 1.36e-05 1015 tok/s
+step 8125/9000 loss 2.5123 lr 1.34e-05 946 tok/s
+step 8150/9000 loss 2.4403 lr 1.32e-05 990 tok/s
+step 8175/9000 loss 2.4637 lr 1.30e-05 1011 tok/s
+step 8200/9000 loss 2.5081 lr 1.28e-05 1012 tok/s
+step 8225/9000 loss 2.3837 lr 1.27e-05 975 tok/s
+step 8250/9000 loss 2.6191 lr 1.25e-05 977 tok/s
+step 8275/9000 loss 2.4923 lr 1.23e-05 893 tok/s
+step 8300/9000 loss 2.6071 lr 1.22e-05 983 tok/s
+step 8325/9000 loss 2.5648 lr 1.20e-05 1030 tok/s
+step 8350/9000 loss 2.4150 lr 1.19e-05 1027 tok/s
+step 8375/9000 loss 2.2823 lr 1.17e-05 1031 tok/s
+step 8400/9000 loss 2.6010 lr 1.16e-05 1038 tok/s
+step 8425/9000 loss 2.7018 lr 1.15e-05 1038 tok/s
+step 8450/9000 loss 2.5901 lr 1.13e-05 1002 tok/s
+step 8475/9000 loss 2.5358 lr 1.12e-05 991 tok/s
+step 8500/9000 loss 2.4637 lr 1.11e-05 1023 tok/s
+  [eval step 8500] val_loss 2.5162 (best 2.5162)
+  [gen] Once upon a time,  in a small small house with a big tree with many toys of friends. It liked to play in the water every day and make things all sorts of games and animals in the garden every day. And everyone would always eat from from one day on when he could take something new and she put it on on her big blue rock and made everyone laugh.
+  [gen] The little girl wanted to  help the girl, so the cat said, "Hi, bird! You are a nice boy who is just a little fish. The cat was a happy friend.
+  [gen] In the dark forest,  there was a big tree. The small car had an enormous red ball in the garden. Every day, the blue sea was very happy to help his friends.
+  saved ckpt/nlp_full/model_8500.pt
+step 8525/9000 loss 2.5972 lr 1.10e-05 642 tok/s
+step 8550/9000 loss 2.5001 lr 1.09e-05 1044 tok/s
+step 8575/9000 loss 2.5291 lr 1.08e-05 1036 tok/s
+step 8600/9000 loss 2.4749 lr 1.07e-05 1039 tok/s
+step 8625/9000 loss 2.3717 lr 1.06e-05 1037 tok/s
+step 8650/9000 loss 2.5911 lr 1.05e-05 1043 tok/s
+step 8675/9000 loss 2.4419 lr 1.05e-05 1041 tok/s
+step 8700/9000 loss 2.5415 lr 1.04e-05 1041 tok/s
+step 8725/9000 loss 2.4459 lr 1.03e-05 1043 tok/s
+step 8750/9000 loss 2.4510 lr 1.03e-05 1038 tok/s
+step 8775/9000 loss 2.3479 lr 1.02e-05 1006 tok/s
+step 8800/9000 loss 2.4827 lr 1.02e-05 1006 tok/s
+step 8825/9000 loss 2.5894 lr 1.01e-05 1030 tok/s
+step 8850/9000 loss 2.5575 lr 1.01e-05 1036 tok/s
+step 8875/9000 loss 2.5774 lr 1.01e-05 1029 tok/s
+step 8900/9000 loss 2.6247 lr 1.00e-05 1030 tok/s
+step 8925/9000 loss 2.6068 lr 1.00e-05 1033 tok/s
+step 8950/9000 loss 2.4513 lr 1.00e-05 1034 tok/s
+step 8975/9000 loss 2.6472 lr 1.00e-05 1036 tok/s
+step 9000/9000 loss 2.6618 lr 1.00e-05 1029 tok/s
+  [eval step 9000] val_loss 2.5092 (best 2.5092)
+  [gen] Once upon a time,  in a small house with his friends, her mom and dad. They played together every day and had lots of fun.
+  [gen] The little girl wanted to  see the little fish. The little boy said, "Hi, little bird. The cat was sad because it could not find his toy.
+  [gen] In the dark forest,  there was a small cat. The bird liked to play with its friends in the yard.
+  saved ckpt/nlp_full/model_9000.pt
+done

logs/train_full_smoke.log

@@ -0,0 +1,18 @@
+resumed ckpt/nlp/model_3200.pt at step 3200
+params: 7.79M  train_tokens: 528,000,852  steps_per_epoch: 128906
+step 3210/3250 loss 3.1068 lr 1.01e-05 533 tok/s
+step 3220/3250 loss 3.1343 lr 1.00e-05 729 tok/s
+  [eval step 3225] val_loss 3.1498 (best 3.0871)
+  [gen] Once upon a time,  there lived in a big garden and they all played games together every day long. They wanted to play with their friends playing with her best friend came by himself.
+  [gen] The little girl wanted to  play with the big tree with his friends, and it was a very happy. One day, there lived in a small garden with many things. The little bee did not know what it could help. But then something unexpected happened! A little boy was so excited and said, "Yes, I'm glad to make you new things can be more careful!" The bird looked at the big, friendly cat
+  [gen] In the dark forest,  the end of the day on a small house. The big trees in the animals, friends went to the shiny thing inside the box. They thanked Lily and the little boat became best friends. The small ball flew and down his friend, a little boy named Whiskers. Spot loved to play with all lived happily together every day. One sunny day, they played together and had so fun at home. One
+  saved ckpt/nlp_full/model_3225.pt
+step 3230/3250 loss 3.1233 lr 1.00e-05 341 tok/s
+step 3240/3250 loss 3.2625 lr 1.00e-05 730 tok/s
+step 3250/3250 loss 3.0625 lr 1.00e-05 766 tok/s
+  [eval step 3250] val_loss 3.1377 (best 3.0871)
+  [gen] Once upon a time,  there was a special day long.
+  [gen] The little girl wanted to  help his friends, the big tree. The cat found a small house near a box on the tree. The man was so happy and played and had fun in a park with the ball. The bird was very excited and wanted to play with it. Tim loved to find the ball. He went to the end of the party and they could take the big red toy too. He saw a small hole
+  [gen] In the dark forest,  there was a very good fun. They had so fun playing in the day on his long stick in the big hill. The people wanted to help his friend to play with it together every day.
+  saved ckpt/nlp_full/model_3250.pt
+done

logs/train_hybrid18m.log

@@ -0,0 +1,274 @@
+resumed ckpt/tiny18m/model_1500.pt at step 1500
+params: 16.77M  train_tokens: 41,369,737  steps_per_epoch: 10100
+step 1525/10000 loss 2.5653 lr 4.82e-05 655 tok/s
+step 1550/10000 loss 2.5523 lr 4.82e-05 633 tok/s
+step 1575/10000 loss 2.4717 lr 4.81e-05 653 tok/s
+step 1600/10000 loss 2.5622 lr 4.80e-05 649 tok/s
+step 1625/10000 loss 2.6742 lr 4.79e-05 654 tok/s
+step 1650/10000 loss 2.7211 lr 4.79e-05 655 tok/s
+step 1675/10000 loss 2.6297 lr 4.78e-05 652 tok/s
+step 1700/10000 loss 2.7394 lr 4.77e-05 650 tok/s
+step 1725/10000 loss 2.6937 lr 4.77e-05 651 tok/s
+step 1750/10000 loss 2.5415 lr 4.76e-05 650 tok/s
+step 1775/10000 loss 2.7357 lr 4.75e-05 652 tok/s
+step 1800/10000 loss 2.4618 lr 4.74e-05 651 tok/s
+step 1825/10000 loss 2.4603 lr 4.73e-05 652 tok/s
+step 1850/10000 loss 2.4765 lr 4.73e-05 652 tok/s
+step 1875/10000 loss 2.5524 lr 4.72e-05 653 tok/s
+step 1900/10000 loss 2.5166 lr 4.71e-05 651 tok/s
+step 1925/10000 loss 2.5092 lr 4.70e-05 654 tok/s
+step 1950/10000 loss 2.6572 lr 4.69e-05 650 tok/s
+step 1975/10000 loss 2.3581 lr 4.68e-05 653 tok/s
+step 2000/10000 loss 2.5298 lr 4.68e-05 652 tok/s
+  [eval step 2000] val_loss 2.5207 (best 2.5207)
+  [gen] Once upon a time,  in a big big green park with many animals. One day, there was a big tree in the snow. It was very shiny and shiny.
+  [gen] The little girl wanted to  visit the little girl. The fish said, "Hello, little boy! Can you help me find my friend?"
+  [gen] In the dark forest,  there lived a little girl named Kitty. Whiskers loved to play with her friends in the park. One day, Bob went out for a walk and find food for dinner.
+  saved ckpt/tiny18m/model_2000.pt
+step 2025/10000 loss 2.3721 lr 4.67e-05 423 tok/s
+step 2050/10000 loss 2.6300 lr 4.66e-05 654 tok/s
+step 2075/10000 loss 2.5421 lr 4.65e-05 654 tok/s
+step 2100/10000 loss 2.6857 lr 4.64e-05 653 tok/s
+step 2125/10000 loss 2.3957 lr 4.63e-05 651 tok/s
+step 2150/10000 loss 2.4925 lr 4.62e-05 652 tok/s
+step 2175/10000 loss 2.4657 lr 4.61e-05 650 tok/s
+step 2200/10000 loss 2.5944 lr 4.60e-05 652 tok/s
+step 2225/10000 loss 2.4952 lr 4.59e-05 652 tok/s
+step 2250/10000 loss 2.5284 lr 4.58e-05 630 tok/s
+step 2275/10000 loss 2.5711 lr 4.57e-05 650 tok/s
+step 2300/10000 loss 2.4540 lr 4.56e-05 653 tok/s
+step 2325/10000 loss 2.6500 lr 4.55e-05 651 tok/s
+step 2350/10000 loss 2.4963 lr 4.54e-05 655 tok/s
+step 2375/10000 loss 2.4619 lr 4.53e-05 653 tok/s
+step 2400/10000 loss 2.6242 lr 4.52e-05 650 tok/s
+step 2425/10000 loss 2.3576 lr 4.51e-05 650 tok/s
+step 2450/10000 loss 2.4192 lr 4.50e-05 645 tok/s
+step 2475/10000 loss 2.6215 lr 4.49e-05 649 tok/s
+step 2500/10000 loss 2.4023 lr 4.48e-05 651 tok/s
+  [eval step 2500] val_loss 2.4974 (best 2.4974)
+  [gen] Once upon a time,  there was a little dog named Spot. loved to play outside all day on the big tree. The ball had many friends in the pond and played a fun game together every day.
+  [gen] The little girl wanted to  help the little girl. The little girl said, "Hi! My name is The big boy. The dog had a friend, a cat. The boy said, "Hello, little bird, and little bird. The rabbit wanted to help. He asked his friends to help the small bug. So, the little bird flew by the tree and asked, "What are you doing? What do you
+  [gen] In the dark forest,  there was a small boy with the red ball in the tree. The bird liked to play near the big house by the tree every day.
+step 2525/10000 loss 2.5237 lr 4.47e-05 404 tok/s
+step 2550/10000 loss 2.6934 lr 4.46e-05 652 tok/s
+step 2575/10000 loss 2.5977 lr 4.45e-05 652 tok/s
+step 2600/10000 loss 2.6915 lr 4.44e-05 649 tok/s
+step 2625/10000 loss 2.7244 lr 4.43e-05 651 tok/s
+step 2650/10000 loss 2.5833 lr 4.41e-05 652 tok/s
+step 2675/10000 loss 2.5070 lr 4.40e-05 651 tok/s
+step 2700/10000 loss 2.6618 lr 4.39e-05 651 tok/s
+step 2725/10000 loss 2.3996 lr 4.38e-05 652 tok/s
+step 2750/10000 loss 2.5017 lr 4.37e-05 650 tok/s
+step 2775/10000 loss 2.5809 lr 4.36e-05 650 tok/s
+step 2800/10000 loss 2.4659 lr 4.34e-05 652 tok/s
+step 2825/10000 loss 2.4142 lr 4.33e-05 642 tok/s
+step 2850/10000 loss 2.6333 lr 4.32e-05 651 tok/s
+step 2875/10000 loss 2.3603 lr 4.31e-05 651 tok/s
+step 2900/10000 loss 2.5381 lr 4.30e-05 648 tok/s
+step 2925/10000 loss 2.4836 lr 4.28e-05 650 tok/s
+step 2950/10000 loss 2.4351 lr 4.27e-05 646 tok/s
+step 2975/10000 loss 2.4030 lr 4.26e-05 652 tok/s
+step 3000/10000 loss 2.5169 lr 4.25e-05 651 tok/s
+  [eval step 3000] val_loss 2.4770 (best 2.4770)
+  [gen] Once upon a time,  there was a little bird. The cat liked to play with the big ball in the water. One day, it came down from behind the tree to see what was inside the house.
+  [gen] The little girl wanted to  help the little boy. The dog asked the little cat, "Can you help me find your friend my friend first." The big bird said, "Hello! I will help you!"
+  [gen] In the dark forest,  there was a little bug. The bird lived in a big tree. All the animals wanted to play on the tree near his home. They flew around and played with their friends.
+  saved ckpt/tiny18m/model_3000.pt
+step 3025/10000 loss 2.4765 lr 4.23e-05 416 tok/s
+step 3050/10000 loss 2.3956 lr 4.22e-05 652 tok/s
+step 3075/10000 loss 2.4885 lr 4.21e-05 661 tok/s
+step 3100/10000 loss 2.4010 lr 4.20e-05 735 tok/s
+step 3125/10000 loss 2.4476 lr 4.18e-05 736 tok/s
+step 3150/10000 loss 2.3956 lr 4.17e-05 733 tok/s
+step 3175/10000 loss 2.3971 lr 4.16e-05 738 tok/s
+step 3200/10000 loss 2.2890 lr 4.14e-05 732 tok/s
+step 3225/10000 loss 2.4471 lr 4.13e-05 734 tok/s
+step 3250/10000 loss 2.3584 lr 4.12e-05 735 tok/s
+step 3275/10000 loss 2.6257 lr 4.10e-05 730 tok/s
+step 3300/10000 loss 2.4169 lr 4.09e-05 734 tok/s
+step 3325/10000 loss 2.4542 lr 4.08e-05 730 tok/s
+step 3350/10000 loss 2.5349 lr 4.06e-05 721 tok/s
+step 3375/10000 loss 2.4588 lr 4.05e-05 723 tok/s
+step 3400/10000 loss 2.5400 lr 4.04e-05 722 tok/s
+step 3425/10000 loss 2.4968 lr 4.02e-05 734 tok/s
+step 3450/10000 loss 2.4677 lr 4.01e-05 737 tok/s
+step 3475/10000 loss 2.6485 lr 4.00e-05 733 tok/s
+step 3500/10000 loss 2.2768 lr 3.98e-05 735 tok/s
+  [eval step 3500] val_loss 2.4624 (best 2.4624)
+  [gen] Once upon a time,  in a small house and many toys. One day, it was very cold outside her room! Lily's toy car had a new toy, so she opened the door with a small pink cup inside. The toy car was not easy to do things like Tim, but they both went to her house to find something special.
+  [gen] The little girl wanted to  help the little boy. The boy said, "I can help you find my way home." So, Bob took the bug home. He found a new friend. But Tom could not eat it.
+  [gen] In the dark forest,  there was a little bird. The bird was very slow and sad too because he lost his toy.
+step 3525/10000 loss 2.3328 lr 3.97e-05 466 tok/s
+step 3550/10000 loss 2.4918 lr 3.95e-05 733 tok/s
+step 3575/10000 loss 2.5129 lr 3.94e-05 697 tok/s
+step 3600/10000 loss 2.6261 lr 3.93e-05 738 tok/s
+step 3625/10000 loss 2.4719 lr 3.91e-05 735 tok/s
+step 3650/10000 loss 2.4097 lr 3.90e-05 735 tok/s
+step 3675/10000 loss 2.5441 lr 3.88e-05 737 tok/s
+step 3700/10000 loss 2.4822 lr 3.87e-05 735 tok/s
+step 3725/10000 loss 2.4271 lr 3.85e-05 728 tok/s
+step 3750/10000 loss 2.3713 lr 3.84e-05 739 tok/s
+step 3775/10000 loss 2.5107 lr 3.82e-05 730 tok/s
+step 3800/10000 loss 2.6896 lr 3.81e-05 736 tok/s
+step 3825/10000 loss 2.3038 lr 3.79e-05 736 tok/s
+step 3850/10000 loss 2.5466 lr 3.78e-05 734 tok/s
+step 3875/10000 loss 2.3960 lr 3.77e-05 731 tok/s
+step 3900/10000 loss 2.3812 lr 3.75e-05 731 tok/s
+step 3925/10000 loss 2.3197 lr 3.74e-05 739 tok/s
+step 3950/10000 loss 2.5414 lr 3.72e-05 738 tok/s
+step 3975/10000 loss 2.7610 lr 3.71e-05 738 tok/s
+step 4000/10000 loss 2.5497 lr 3.69e-05 740 tok/s
+  [eval step 4000] val_loss 2.4493 (best 2.4493)
+  [gen] Once upon a time,  in a small house near the big tree to play in the park. They had so much fun together.
+  [gen] The little girl wanted to  help the bird. The big bug was sad. But, the cat saw a little dog playing with the ball.
+  [gen] In the dark forest,  a little girl went to the park with many yummy things. They all liked them very much. The dog liked to play together every day. One day, they saw a big red ball in the pond. It was an idea. But then, there was a big red car on the ground. The tree had many pretty colors and thought it would be fun to play.
+  saved ckpt/tiny18m/model_4000.pt
+step 4025/10000 loss 2.5246 lr 3.68e-05 465 tok/s
+step 4050/10000 loss 2.5094 lr 3.66e-05 740 tok/s
+step 4075/10000 loss 2.4757 lr 3.65e-05 740 tok/s
+step 4100/10000 loss 2.5315 lr 3.63e-05 736 tok/s
+step 4125/10000 loss 2.4815 lr 3.61e-05 738 tok/s
+step 4150/10000 loss 2.3388 lr 3.60e-05 734 tok/s
+step 4175/10000 loss 2.5207 lr 3.58e-05 694 tok/s
+step 4200/10000 loss 2.4453 lr 3.57e-05 639 tok/s
+step 4225/10000 loss 2.3440 lr 3.55e-05 650 tok/s
+step 4250/10000 loss 2.5933 lr 3.54e-05 651 tok/s
+step 4275/10000 loss 2.5590 lr 3.52e-05 649 tok/s
+step 4300/10000 loss 2.5445 lr 3.51e-05 651 tok/s
+step 4325/10000 loss 2.5882 lr 3.49e-05 650 tok/s
+step 4350/10000 loss 2.3221 lr 3.48e-05 655 tok/s
+step 4375/10000 loss 2.5339 lr 3.46e-05 654 tok/s
+step 4400/10000 loss 2.5916 lr 3.45e-05 654 tok/s
+step 4425/10000 loss 2.7155 lr 3.43e-05 654 tok/s
+step 4450/10000 loss 2.3446 lr 3.41e-05 650 tok/s
+step 4475/10000 loss 2.5560 lr 3.40e-05 650 tok/s
+step 4500/10000 loss 2.3556 lr 3.38e-05 651 tok/s
+  [eval step 4500] val_loss 2.4327 (best 2.4327)
+  [gen] Once upon a time,  there was a little bird named Fin. Bob loved to play with his ball every day long. One sunny day, while playing outside, there was an adventurous cat. He loved to look at the tree every day and run around in it.
+  [gen] The little girl wanted to  help the cat. The mouse said, "Can I play with you?" The little bird said, "I am a naughty dog. Let's see what was in the tree. The bird asked, "Can you show me my friend the bird."
+  [gen] In the dark forest,  there lived a little girl named Kitty. She loved to play outside with her friends in the park. One day, Lucy saw a small girl who was playing near it. He walked up and saw a cat sitting on a branch that said "Hello! Can I have some food?"
+step 4525/10000 loss 2.5199 lr 3.37e-05 405 tok/s
+step 4550/10000 loss 2.4112 lr 3.35e-05 649 tok/s
+step 4575/10000 loss 2.5137 lr 3.34e-05 651 tok/s
+step 4600/10000 loss 2.2691 lr 3.32e-05 649 tok/s
+step 4625/10000 loss 2.5456 lr 3.30e-05 643 tok/s
+step 4650/10000 loss 2.6766 lr 3.29e-05 650 tok/s
+step 4675/10000 loss 2.5706 lr 3.27e-05 650 tok/s
+step 4700/10000 loss 2.4921 lr 3.26e-05 653 tok/s
+step 4725/10000 loss 2.6044 lr 3.24e-05 653 tok/s
+step 4750/10000 loss 2.4704 lr 3.22e-05 653 tok/s
+step 4775/10000 loss 2.5191 lr 3.21e-05 650 tok/s
+step 4800/10000 loss 2.2880 lr 3.19e-05 652 tok/s
+step 4825/10000 loss 2.6277 lr 3.18e-05 652 tok/s
+step 4850/10000 loss 2.3137 lr 3.16e-05 649 tok/s
+step 4875/10000 loss 2.4127 lr 3.14e-05 649 tok/s
+step 4900/10000 loss 6.7132 lr 3.13e-05 654 tok/s
+step 4925/10000 loss 5.6766 lr 3.11e-05 653 tok/s
+step 4950/10000 loss 5.1870 lr 3.10e-05 650 tok/s
+step 4975/10000 loss 4.7262 lr 3.08e-05 656 tok/s
+step 5000/10000 loss 4.2589 lr 3.06e-05 652 tok/s
+  [eval step 5000] val_loss 2.8255 (best 2.4327)
+  [gen] Once upon a time,  there was a big, in this big tree, there lived an there, on the other birds. One sunny day, there was an an an by that in the room in one room, in it, there was red and a many fun things inside.
+  [gen] The little girl wanted to  help the little girl, can't fly in the tree. The green bird was sad but lost it had come back to play with others.
+  [gen] In the dark forest,  there was a little red ball that was so high in the sky. It lived near the tree and wanted to go on it. So, he went outside with his friend, they got a new place for playing with a big party. One day, while Spot was in the tree, he found many pretty toys and things and run around with other friends. They played all day long and had lots of
+  saved ckpt/tiny18m/model_5000.pt
+step 5025/10000 loss 4.0110 lr 3.05e-05 400 tok/s
+step 5050/10000 loss 3.0077 lr 3.03e-05 653 tok/s
+step 5075/10000 loss 3.3599 lr 3.02e-05 652 tok/s
+step 5100/10000 loss 3.1507 lr 3.00e-05 655 tok/s
+step 5125/10000 loss 3.0741 lr 2.98e-05 650 tok/s
+step 5150/10000 loss 2.6533 lr 2.97e-05 652 tok/s
+step 5175/10000 loss 2.5485 lr 2.95e-05 651 tok/s
+step 5200/10000 loss 2.9739 lr 2.94e-05 648 tok/s
+step 5225/10000 loss 2.9358 lr 2.92e-05 654 tok/s
+step 5250/10000 loss 2.9247 lr 2.90e-05 651 tok/s
+step 5275/10000 loss 3.0765 lr 2.89e-05 653 tok/s
+step 5300/10000 loss 2.6095 lr 2.87e-05 651 tok/s
+step 5325/10000 loss 2.3876 lr 2.86e-05 651 tok/s
+step 5350/10000 loss 1.9379 lr 2.84e-05 651 tok/s
+step 5375/10000 loss 6.1126 lr 2.82e-05 650 tok/s
+step 5400/10000 loss 5.6536 lr 2.81e-05 651 tok/s
+step 5425/10000 loss 5.1948 lr 2.79e-05 653 tok/s
+step 5450/10000 loss 5.3294 lr 2.78e-05 651 tok/s
+step 5475/10000 loss 5.3229 lr 2.76e-05 654 tok/s
+step 5500/10000 loss 5.3180 lr 2.74e-05 652 tok/s
+  [eval step 5500] val_loss 3.5666 (best 2.4327)
+  [gen] Once upon a time,  a little house with the warm blue, red car, and the car in the ball, they would not fly high high in the sun.
+  [gen] The little girl wanted to  eat the food from the red ball and the new toys, and, but they could not.
+  [gen] In the dark forest,  the new friend, the big tree in a tree, and the other big boat, the big red car, green green and the all of the toys, red ball, thats, the toy is not a place to be strong and strong because you can't play with the big cat.
+step 5525/10000 loss 4.8390 lr 2.73e-05 417 tok/s
+step 5550/10000 loss 4.8588 lr 2.71e-05 654 tok/s
+step 5575/10000 loss 4.9937 lr 2.70e-05 652 tok/s
+step 5600/10000 loss 4.8477 lr 2.68e-05 653 tok/s
+step 5625/10000 loss 4.3689 lr 2.66e-05 653 tok/s
+step 5650/10000 loss 4.3986 lr 2.65e-05 655 tok/s
+step 5675/10000 loss 3.2021 lr 2.63e-05 653 tok/s
+step 5700/10000 loss 4.9449 lr 2.62e-05 654 tok/s
+step 5725/10000 loss 4.9584 lr 2.60e-05 700 tok/s
+step 5750/10000 loss 4.4951 lr 2.59e-05 736 tok/s
+step 5775/10000 loss 4.5459 lr 2.57e-05 741 tok/s
+step 5800/10000 loss 4.8274 lr 2.55e-05 740 tok/s
+step 5825/10000 loss 4.4094 lr 2.54e-05 735 tok/s
+step 5850/10000 loss 4.8188 lr 2.52e-05 742 tok/s
+step 5875/10000 loss 4.3918 lr 2.51e-05 744 tok/s
+step 5900/10000 loss 4.6424 lr 2.49e-05 728 tok/s
+step 5925/10000 loss 4.3456 lr 2.48e-05 679 tok/s
+step 5950/10000 loss 4.6885 lr 2.46e-05 651 tok/s
+step 5975/10000 loss 4.4286 lr 2.45e-05 654 tok/s
+step 6000/10000 loss 4.4459 lr 2.43e-05 655 tok/s
+  [eval step 6000] val_loss 3.9556 (best 2.4327)
+  [gen] Once upon a time,  in the A small red ball and go on its way to the tree.
+  [gen] The little girl wanted to  the nest to the tree. The door will not be a long way into it, it was the rest of this fun things and play with it all of the high red ball in its hand all day, but you can make it with it all all new places it's when he need, he could use that it will be a special place, or
+  [gen] In the dark forest,  the sun, the small in the tree, all of the the other other toys, and the other kids to the other friends, and more fun things, and they would not not find any way to the other place, the best for one last, and they could always, if you can, red, one, the
+  saved ckpt/tiny18m/model_6000.pt
+step 6025/10000 loss 3.8443 lr 2.42e-05 403 tok/s
+step 6050/10000 loss 4.0320 lr 2.40e-05 650 tok/s
+step 6075/10000 loss 4.1920 lr 2.39e-05 652 tok/s
+step 6100/10000 loss 4.3014 lr 2.37e-05 652 tok/s
+step 6125/10000 loss 4.5500 lr 2.35e-05 653 tok/s
+step 6150/10000 loss 4.1831 lr 2.34e-05 654 tok/s
+step 6175/10000 loss 4.2021 lr 2.32e-05 651 tok/s
+step 6200/10000 loss 3.9771 lr 2.31e-05 650 tok/s
+step 6225/10000 loss 3.6250 lr 2.29e-05 654 tok/s
+step 6250/10000 loss 4.0360 lr 2.28e-05 648 tok/s
+step 6275/10000 loss 3.6614 lr 2.26e-05 650 tok/s
+step 6300/10000 loss 4.0606 lr 2.25e-05 645 tok/s
+step 6325/10000 loss 3.6680 lr 2.23e-05 650 tok/s
+step 6350/10000 loss 4.3962 lr 2.22e-05 649 tok/s
+step 6375/10000 loss 4.0765 lr 2.21e-05 651 tok/s
+step 6400/10000 loss 3.5979 lr 2.19e-05 654 tok/s
+step 6425/10000 loss 3.6755 lr 2.18e-05 651 tok/s
+step 6450/10000 loss 3.9469 lr 2.16e-05 651 tok/s
+step 6475/10000 loss 4.0041 lr 2.15e-05 651 tok/s
+step 6500/10000 loss 4.1386 lr 2.13e-05 647 tok/s
+  [eval step 6500] val_loss 4.1906 (best 2.4327)
+  [gen] Once upon a time,  there was a big, soft, big, soft soft soft, green ball and red ball with the ball each a long red ball it in the tree and into the high high high up with its red ball ball into the shiny toy car all all balls all all toys and it and all one would all all other toys as best friends, but the high tree best friend nest back on his pretty soft soft
+  [gen] The little girl wanted to  the people in the tree.
+  [gen] In the dark forest,  the other one more fun things from the other things in the tree and in its softing, but no longer the box, all the other other food, in the big toy can, and the pretty rocks, the red, red,,, and in the tree, the cat and the people can be in the Inm this tree, the other of the end of the other other things,,
+step 6525/10000 loss 3.9395 lr 2.12e-05 392 tok/s
+step 6550/10000 loss 3.8221 lr 2.10e-05 657 tok/s
+step 6575/10000 loss 3.9042 lr 2.09e-05 655 tok/s
+step 6600/10000 loss 3.6134 lr 2.07e-05 655 tok/s
+step 6625/10000 loss 3.5598 lr 2.06e-05 652 tok/s
+step 6650/10000 loss 4.2290 lr 2.05e-05 658 tok/s
+step 6675/10000 loss 3.9047 lr 2.03e-05 652 tok/s
+step 6700/10000 loss 3.7798 lr 2.02e-05 648 tok/s
+step 6725/10000 loss 4.3956 lr 2.00e-05 651 tok/s
+step 6750/10000 loss 4.0952 lr 1.99e-05 652 tok/s
+step 6775/10000 loss 3.9629 lr 1.98e-05 652 tok/s
+step 6800/10000 loss 4.1737 lr 1.96e-05 654 tok/s
+step 6825/10000 loss 3.5167 lr 1.95e-05 653 tok/s
+step 6850/10000 loss 4.1574 lr 1.94e-05 653 tok/s
+step 6875/10000 loss 3.9978 lr 1.92e-05 655 tok/s
+step 6900/10000 loss 3.6815 lr 1.91e-05 688 tok/s
+step 6925/10000 loss 3.5866 lr 1.90e-05 735 tok/s
+step 6950/10000 loss 3.7959 lr 1.88e-05 736 tok/s
+step 6975/10000 loss 3.6071 lr 1.87e-05 737 tok/s
+step 7000/10000 loss 4.5193 lr 1.86e-05 742 tok/s
+  [eval step 7000] val_loss 4.4284 (best 2.4327)
+  [gen] Once upon a time,  there was a little redtion in the tree with the big yellow nest in it, and there was the ball all all its friends best friends, and he all the new---- of the newicense, but you can use this file, there's what would if we start the key, or the other side of the Lics to other."
+  [gen] The little girl wanted to  the All of the other dogs more people can be fun things in the tree all more leaves from the soft, fun, and it will have a new, and any new world is on its best job, they are all all all both both the new way with the your long red, red, softlub, or
+  [gen] In the dark forest,  the tree all other.  It can be used to, it will be useful, and you work with a new red or, red This, and the
+  saved ckpt/tiny18m/model_7000.pt
+step 7025/10000 loss 4.0391 lr 1.84e-05 449 tok/s
+step 7050/10000 loss 3.7119 lr 1.83e-05 740 tok/s

logs/train_judge.log

@@ -0,0 +1,22 @@
+train 237 val 39 weights [3.0, 5.925000190734863, 4.647058963775635, 11.285714149475098, 5.15217399597168]
+epoch 1/20 val_acc 0.179 per-class {'true statement': 0.0, 'false statement': 1.0, 'supports': 0.0, 'refutes': 0.333, 'not_enough_info': 0.0} (13s)
+epoch 2/20 val_acc 0.205 per-class {'true statement': 0.0, 'false statement': 1.0, 'supports': 0.111, 'refutes': 0.333, 'not_enough_info': 0.0} (14s)
+epoch 3/20 val_acc 0.231 per-class {'true statement': 0.0, 'false statement': 1.0, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.0} (12s)
+epoch 4/20 val_acc 0.256 per-class {'true statement': 0.0, 'false statement': 1.0, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (12s)
+epoch 5/20 val_acc 0.205 per-class {'true statement': 0.0, 'false statement': 0.833, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.125} (13s)
+epoch 6/20 val_acc 0.205 per-class {'true statement': 0.0, 'false statement': 0.833, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.125} (13s)
+epoch 7/20 val_acc 0.256 per-class {'true statement': 0.077, 'false statement': 0.833, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (13s)
+epoch 8/20 val_acc 0.231 per-class {'true statement': 0.077, 'false statement': 0.667, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (12s)
+epoch 9/20 val_acc 0.231 per-class {'true statement': 0.154, 'false statement': 0.5, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (13s)
+epoch 10/20 val_acc 0.256 per-class {'true statement': 0.231, 'false statement': 0.5, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (14s)
+epoch 11/20 val_acc 0.231 per-class {'true statement': 0.231, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (13s)
+epoch 12/20 val_acc 0.256 per-class {'true statement': 0.231, 'false statement': 0.5, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.125} (13s)
+epoch 13/20 val_acc 0.256 per-class {'true statement': 0.231, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.25} (13s)
+epoch 14/20 val_acc 0.282 per-class {'true statement': 0.231, 'false statement': 0.5, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.25} (13s)
+epoch 15/20 val_acc 0.282 per-class {'true statement': 0.231, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.375} (12s)
+epoch 16/20 val_acc 0.282 per-class {'true statement': 0.308, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.333, 'not_enough_info': 0.25} (13s)
+epoch 17/20 val_acc 0.308 per-class {'true statement': 0.385, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.375} (13s)
+epoch 18/20 val_acc 0.308 per-class {'true statement': 0.385, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.375} (13s)
+epoch 19/20 val_acc 0.333 per-class {'true statement': 0.462, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.375} (13s)
+epoch 20/20 val_acc 0.282 per-class {'true statement': 0.385, 'false statement': 0.333, 'supports': 0.222, 'refutes': 0.0, 'not_enough_info': 0.25} (12s)
+done -> ckpt/judge/judge.pt best_acc 0.333

logs/train_lora_i.log

@@ -0,0 +1,51 @@
+base model_best.pt | lora adapters 56 | trainable 1,156,032
+train 827 eval 75 filtered 0
+step 25/412 loss 3.0658 sft 2.9139 143 tok/s
+  [eval 25] sft_val_loss 3.2996 val_ppl 8.46 [best ppl] [new best]
+    sample: Check the claim against the record. Verdict: true. Confidence: MEvaluption: Newsignormary forcement; this claim is
+step 50/412 loss 4.0365 sft 3.9101 122 tok/s
+  [eval 50] sft_val_loss 3.0890 val_ppl 10.71 [new best]
+    sample: Step 209 -> the average Statement is distributed in this stimitions and Rublished records, so the claim reports conflict on the year, with the 25
+step 75/412 loss 3.0832 sft 2.9491 133 tok/s
+  [eval 75] sft_val_loss 2.9774 val_ppl 10.45 [new best]
+    sample: Step 209-mail (Context: a speech in Newsging Publiciated in the Statement, and I am not average; I can't read
+step 100/412 loss 2.0886 sft 1.9155 136 tok/s
+  [eval 100] sft_val_loss 2.9385 val_ppl 15.74 [new best]
+    sample: Stepartment date"
+step 125/412 loss 2.3968 sft 2.2467 145 tok/s
+  [eval 125] sft_val_loss 2.8948 val_ppl 12.03 [new best]
+    sample: Claims (2) the records agree and the claim is not established records; no independent Verdict: refutes. Confidence: ME
+step 150/412 loss 2.5207 sft 2.3751 121 tok/s
+  [eval 150] sft_val_loss 2.8605 val_ppl 19.63 [new best]
+    sample: Claims addression 3 of the 20000-39 - Prolled in the Statementment (2) Mitteech to the Stepospicle's author
+step 175/412 loss 3.8445 sft 3.7116 134 tok/s
+  [eval 175] sft_val_loss 2.8434 val_ppl 18.96 [new best]
+    sample: Claims (
+step 200/412 loss 2.7964 sft 2.6536 155 tok/s
+  [eval 200] sft_val_loss 2.8332 val_ppl 9.73 [new best]
+    sample: Step 2015; (2) the claim is not a signalysment in this statement reports 3,0000 totror identified the record, and a
+step 225/412 loss 2.7373 sft 2.6080 146 tok/s
+  [eval 225] sft_val_loss 2.8124 val_ppl 9.12 [new best]
+    sample: Claim's 20986; (25) 3 of the 24997,0003,00 million 307-check: 55539999
+step 250/412 loss 2.9244 sft 2.7815 146 tok/s
+  [eval 250] sft_val_loss 2.7948 val_ppl 12.33 [new best]
+    sample: Claim: the 45 million in a Statement: an units to 98; (2) 20,0000 -17% and 3. Sevem: 1
+step 275/412 loss 3.0943 sft 2.9528 145 tok/s
+  [eval 275] sft_val_loss 2.7764 val_ppl 14.68 [new best]
+    sample: Checklist:
+step 300/412 loss 3.1095 sft 2.9730 156 tok/s
+  [eval 300] sft_val_loss 2.7581 val_ppl 27.96 [new best]
+    sample: Claim: the (1,0); 45 million 2978,2000. Stepartment (3) authority reports 3 totribution by
+step 325/412 loss 2.9545 sft 2.8073 146 tok/s
+  [eval 325] sft_val_loss 2.7442 val_ppl 14.99 [new best]
+    sample: Claim: the two period and the programs, a 300 -45-moe. Step 2 record: (3). Stepartment report: the 30
+step 350/412 loss 2.7560 sft 2.6103 146 tok/s
+  [eval 350] sft_val_loss 2.7338 val_ppl 15.39 [new best]
+    sample: Step 2015 -9 million $287; the Stepublication reports 3,000484 but Wiscurate Prollymary COP
+step 375/412 loss 2.0681 sft 1.9116 146 tok/s
+  [eval 375] sft_val_loss 2.7198 val_ppl 17.81 [new best]
+    sample: Claim's the 405 million 2897 - Tex Steposits 3 of the two years. Step 2 record: the Presential debate reports the $2
+step 400/412 loss 3.0132 sft 2.8776 143 tok/s
+  [eval 400] sft_val_loss 2.7207 val_ppl 25.96
+    sample: The 30-3 million 255,295 -45 (C) 28766 $1199985
+done -> ckpt/tiny25m_lora_i best_sft=2.7198 best_ppl=8.46

logs/train_lora_ii.log

@@ -0,0 +1,52 @@
+base model_best.pt | lora adapters 56 | trainable 1,156,032
+replay: 827 raw items from data/train_phase2b.bin (mixture ratio 0.50)
+train 1654 eval 75 filtered 0
+step 25/413 loss 3.1495 sft 2.9957 160 tok/s
+  [eval 25] sft_val_loss 3.8395 val_ppl 8.73 [best ppl] [new best]
+    sample: Checklistry: identifiable entity or event; it requires a source beyond the claim itself. They were amazed by the statementmentmentment indicates to the same year old,
+step 50/413 loss 2.1690 sft 2.0016 138 tok/s
+  [eval 50] sft_val_loss 3.4687 val_ppl 10.09 [new best]
+    sample: Check the claim against the record. Verdict: refutes, or fails to address. She saw a big, friendly bear! He was walking towards it and ran towards them.
+step 75/413 loss 3.4305 sft 3.3153 118 tok/s
+  [eval 75] sft_val_loss 3.3641 val_ppl 10.11 [new best]
+    sample: Checklist: ("Context: a speech in the License.
+step 100/413 loss 3.3717 sft 3.2197 123 tok/s
+  [eval 100] sft_val_loss 3.2918 val_ppl 15.06 [new best]
+    sample: Checklist: ("What's the License.
+step 125/413 loss 2.7740 sft 2.6628 146 tok/s
+  [eval 125] sft_val_loss 3.2125 val_ppl 10.00 [new best]
+    sample: Checklistist: ("What's the Repartment diverges from established records; no independent verified the claim itself. Verdict: true st
+step 150/413 loss 3.3583 sft 3.1279 143 tok/s
+  [eval 150] sft_val_loss 3.1442 val_ppl 16.81 [new best]
+    sample: Checklist: ("Neval Python Republicense for more details. He wanted to record it was a picture of the United States. Context: an
+step 175/413 loss 2.6380 sft 2.5172 142 tok/s
+  [eval 175] sft_val_loss 3.1207 val_ppl 17.20 [new best]
+    sample: Checklist: ("Proview on CONDIUM. Reasoning: 209465,23 million to 10745,255 million 3
+step 200/413 loss 2.7076 sft 2.5983 116 tok/s
+  [eval 200] sft_val_loss 3.1024 val_ppl 8.11 [best ppl] [new best]
+    sample: Checklist: identify the factual assertion; it requires a source beyond the claim itself. Verdict: true statement. Confidence:
+step 225/413 loss 2.5819 sft 2.4518 118 tok/s
+  [eval 225] sft_val_loss 3.0873 val_ppl 7.54 [best ppl] [new best]
+    sample: Checklist: (1)
+step 250/413 loss 2.8106 sft 2.6594 125 tok/s
+  [eval 250] sft_val_loss 3.0514 val_ppl 9.87 [new best]
+    sample: Checklist: (
+step 275/413 loss 3.0073 sft 2.8516 126 tok/s
+  [eval 275] sft_val_loss 3.0365 val_ppl 11.09 [new best]
+    sample: Checklist: ("What's that?"
+step 300/413 loss 2.7067 sft 2.5654 125 tok/s
+  [eval 300] sft_val_loss 3.0256 val_ppl 20.87 [new best]
+    sample: Checklist: (1)
+step 325/413 loss 2.4705 sft 2.3342 130 tok/s
+  [eval 325] sft_val_loss 3.0041 val_ppl 10.74 [new best]
+    sample: Checklist: ("What's the same event.' Two accession.
+step 350/413 loss 2.5596 sft 2.4332 124 tok/s
+  [eval 350] sft_val_loss 3.0434 val_ppl 11.84
+    sample: Checklist: (")
+step 375/413 loss 3.1477 sft 2.9876 126 tok/s
+  [eval 375] sft_val_loss 2.9824 val_ppl 14.02 [new best]
+    sample: Compare the claim to the record. the inspection report shows 950, so the accounts conflict on the time they could see the beautiful things. They were very excited to see a big hill
+step 400/413 loss 2.3532 sft 2.1497 117 tok/s
+  [eval 400] sft_val_loss 2.9765 val_ppl 18.78 [new best]
+    sample: Checklist:
+done -> ckpt/tiny25m_lora_ii best_sft=2.9765 best_ppl=7.54

logs/train_phase2b.log

@@ -0,0 +1,76 @@
+init-from ckpt/tiny18m_reset/model_final.pt (weights only, fresh schedule)
+params: 16.77M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 1/8500 loss 3.9323 lr 1.00e-04 16719 tok/s
+step 25/8500 loss 3.0930 lr 1.00e-04 758 tok/s
+step 50/8500 loss 2.7952 lr 1.00e-04 712 tok/s
+step 75/8500 loss 3.4672 lr 1.00e-04 731 tok/s
+step 100/8500 loss 3.4318 lr 1.00e-04 731 tok/s
+step 125/8500 loss 2.6947 lr 1.00e-04 732 tok/s
+step 150/8500 loss 3.2302 lr 1.00e-04 732 tok/s
+step 175/8500 loss 3.1624 lr 1.00e-04 732 tok/s
+step 200/8500 loss 3.2314 lr 1.00e-04 730 tok/s
+step 225/8500 loss 3.0450 lr 1.00e-04 732 tok/s
+step 250/8500 loss 3.0558 lr 1.00e-04 756 tok/s
+step 275/8500 loss 2.8338 lr 1.00e-04 763 tok/s
+step 300/8500 loss 3.1439 lr 1.00e-04 752 tok/s
+step 325/8500 loss 2.8984 lr 9.99e-05 755 tok/s
+step 350/8500 loss 2.6378 lr 9.99e-05 756 tok/s
+step 375/8500 loss 3.1328 lr 9.99e-05 750 tok/s
+step 400/8500 loss 2.8833 lr 9.99e-05 754 tok/s
+step 425/8500 loss 2.9991 lr 9.98e-05 754 tok/s
+step 450/8500 loss 3.0287 lr 9.98e-05 762 tok/s
+step 475/8500 loss 3.0296 lr 9.98e-05 758 tok/s
+step 500/8500 loss 2.6051 lr 9.97e-05 754 tok/s
+  [eval step 500] val_loss 2.5142 (best 2.5142)
+  [gen] Once upon a time,  there was a big, tall tree. He lived in a dry cave and started to draw on the ground with many small bugs. One day, they played with the ball all day long.
+  [gen] The little girl wanted to  help the little girl get the cat out of a big hole in the garden. A little boy went up under a tree and found a small to play with it. The cat loved to play with the toy car all day long.
+  [gen] In the dark forest,  a little girl and her mom looked at the window and saw a tiny door filled with colorful leaves on the ground. There were toys, like lots of animals, a beautiful rainbow rainbow, and a blue flower, and many trees. One day, she saw some baby bird in a house and decided to take it near the tree and decided to play there there with her toys.
+step 525/8500 loss 2.6642 lr 9.97e-05 463 tok/s
+step 550/8500 loss 2.7348 lr 9.96e-05 743 tok/s
+step 575/8500 loss 2.4240 lr 9.95e-05 741 tok/s
+step 600/8500 loss 2.6232 lr 9.95e-05 742 tok/s
+step 625/8500 loss 2.7197 lr 9.94e-05 727 tok/s
+step 650/8500 loss 2.5819 lr 9.93e-05 732 tok/s
+step 675/8500 loss 2.8008 lr 9.93e-05 732 tok/s
+step 700/8500 loss 2.7696 lr 9.92e-05 736 tok/s
+step 725/8500 loss 2.7050 lr 9.91e-05 735 tok/s
+step 750/8500 loss 2.6851 lr 9.90e-05 734 tok/s
+step 775/8500 loss 3.0217 lr 9.89e-05 738 tok/s
+step 800/8500 loss 2.8280 lr 9.88e-05 735 tok/s
+step 825/8500 loss 2.1964 lr 9.87e-05 736 tok/s
+step 850/8500 loss 2.6617 lr 9.86e-05 726 tok/s
+step 875/8500 loss 2.4887 lr 9.85e-05 751 tok/s
+step 900/8500 loss 2.4185 lr 9.84e-05 754 tok/s
+step 925/8500 loss 2.6803 lr 9.83e-05 717 tok/s
+step 950/8500 loss 2.7594 lr 9.82e-05 741 tok/s
+step 975/8500 loss 2.8337 lr 9.81e-05 739 tok/s
+step 1000/8500 loss 2.4138 lr 9.80e-05 720 tok/s
+  [eval step 1000] val_loss 2.4865 (best 2.4865)
+  [gen] Once upon a time,  there was a little red bird. The bug loved to play with his friends in the yard and eat fun things in the forest.
+  [gen] The little girl wanted to  help the little boy. The cat liked how big the lion could talk. The man's friends were happy with the little dog and the cat all the animals.
+  [gen] In the dark forest,  there were two friends, and were playing in the park. They liked to play and laugh together every day. The fish was very good at the race, and they played games for all the time played together.
+  saved ckpt/tiny18m2/model_1000.pt
+step 1025/8500 loss 2.6435 lr 9.78e-05 468 tok/s
+step 1050/8500 loss 2.3508 lr 9.77e-05 706 tok/s
+step 1075/8500 loss 2.7599 lr 9.76e-05 713 tok/s
+step 1100/8500 loss 2.2291 lr 9.74e-05 709 tok/s
+step 1125/8500 loss 2.9427 lr 9.73e-05 710 tok/s
+step 1150/8500 loss 2.9201 lr 9.71e-05 704 tok/s
+step 1175/8500 loss 2.8660 lr 9.70e-05 705 tok/s
+step 1200/8500 loss 2.6601 lr 9.68e-05 707 tok/s
+step 1225/8500 loss 2.6021 lr 9.67e-05 707 tok/s
+step 1250/8500 loss 2.4595 lr 9.65e-05 709 tok/s
+step 1275/8500 loss 2.5738 lr 9.63e-05 706 tok/s
+step 1300/8500 loss 2.7455 lr 9.62e-05 706 tok/s
+step 1325/8500 loss 2.6901 lr 9.60e-05 710 tok/s
+step 1350/8500 loss 2.6673 lr 9.58e-05 702 tok/s
+step 1375/8500 loss 2.6843 lr 9.56e-05 708 tok/s
+step 1400/8500 loss 2.6751 lr 9.54e-05 714 tok/s
+step 1425/8500 loss 2.4620 lr 9.52e-05 716 tok/s
+step 1450/8500 loss 2.6474 lr 9.51e-05 703 tok/s
+step 1475/8500 loss 2.7279 lr 9.49e-05 709 tok/s
+step 1500/8500 loss 2.7333 lr 9.47e-05 724 tok/s
+  [eval step 1500] val_loss 2.4714 (best 2.4714)
+  [gen] Once upon a time,  in a small house lived on a big farm with many friends. The fish had many friends who liked to play together every day long. One day, the car went on an adventure to play near a big blue sea. It wanted to be his best friend to help people who love him to play and have fun together.
+  [gen] The little girl wanted to  help the girl. The big cat gave the owl some food for the bunny all the food. The old man and the dog were happy. They all played together in the small town.
+  [gen] In the dark forest,  there was a small, expensive rock who could play with its friends in the garden. One day, all of the fun toys decided to do things that played with them for more food to each other to find something special to make people happy and having lots of fun.

logs/train_phase2b_g25.log

@@ -0,0 +1,115 @@
+init-from ckpt/tiny25m_grown/model_final.pt (weights only, fresh schedule)
+params: 25.43M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 1/5000 loss 2.3556 lr 1.00e-04 9775 tok/s
+step 25/5000 loss 2.4194 lr 1.00e-04 440 tok/s
+step 50/5000 loss 2.5145 lr 1.00e-04 425 tok/s
+step 75/5000 loss 2.2660 lr 1.00e-04 424 tok/s
+step 100/5000 loss 2.3939 lr 1.00e-04 425 tok/s
+step 125/5000 loss 2.2330 lr 1.00e-04 423 tok/s
+step 150/5000 loss 2.4469 lr 1.00e-04 428 tok/s
+step 175/5000 loss 2.1646 lr 1.00e-04 428 tok/s
+step 200/5000 loss 2.6688 lr 1.00e-04 427 tok/s
+step 225/5000 loss 2.3660 lr 1.00e-04 428 tok/s
+step 250/5000 loss 2.3601 lr 1.00e-04 429 tok/s
+step 275/5000 loss 2.3920 lr 1.00e-04 426 tok/s
+step 300/5000 loss 2.3319 lr 1.00e-04 421 tok/s
+step 325/5000 loss 2.3072 lr 1.00e-04 427 tok/s
+step 350/5000 loss 2.1679 lr 1.00e-04 428 tok/s
+step 375/5000 loss 2.5859 lr 1.00e-04 428 tok/s
+step 400/5000 loss 2.4134 lr 1.00e-04 426 tok/s
+step 425/5000 loss 2.2913 lr 1.00e-04 429 tok/s
+step 450/5000 loss 2.4803 lr 1.00e-04 427 tok/s
+step 475/5000 loss 2.5224 lr 9.99e-05 428 tok/s
+step 500/5000 loss 2.2127 lr 9.99e-05 429 tok/s
+  [eval step 500] val_loss 2.3559 (best 2.3559)
+  [gen] Once upon a time,  there was a little bird who loved to sit and play with his friends.
+  [gen] The little girl wanted to  help the cat with its wing. The bird was sad because it had a big surprise in its hand.
+  [gen] In the dark forest,  there was a small, strong bear and the bear saw some apples, a big, green tree, a garden, a house, and a pond, some more trees, the small houses, a big sky, a red, blue yellow, blue, green green, and green trees, yellow flowers, yellow, and blue stars. It looks in the open, there it, the window and in the
+step 525/5000 loss 2.0616 lr 9.98e-05 248 tok/s
+step 550/5000 loss 2.2812 lr 9.98e-05 429 tok/s
+step 575/5000 loss 2.1708 lr 9.97e-05 428 tok/s
+step 600/5000 loss 2.3612 lr 9.96e-05 426 tok/s
+step 625/5000 loss 2.4473 lr 9.95e-05 425 tok/s
+step 650/5000 loss 2.3191 lr 9.93e-05 426 tok/s
+step 675/5000 loss 2.3447 lr 9.92e-05 428 tok/s
+step 700/5000 loss 2.3567 lr 9.91e-05 424 tok/s
+step 725/5000 loss 2.4185 lr 9.89e-05 425 tok/s
+step 750/5000 loss 2.3730 lr 9.87e-05 428 tok/s
+step 775/5000 loss 2.6297 lr 9.85e-05 428 tok/s
+step 800/5000 loss 2.4307 lr 9.83e-05 428 tok/s
+step 825/5000 loss 1.9741 lr 9.81e-05 429 tok/s
+step 850/5000 loss 2.4251 lr 9.79e-05 428 tok/s
+step 875/5000 loss 2.2014 lr 9.77e-05 426 tok/s
+step 900/5000 loss 2.2655 lr 9.74e-05 420 tok/s
+step 925/5000 loss 2.2274 lr 9.71e-05 425 tok/s
+step 950/5000 loss 2.3119 lr 9.69e-05 427 tok/s
+step 975/5000 loss 2.4769 lr 9.66e-05 424 tok/s
+step 1000/5000 loss 2.1757 lr 9.63e-05 427 tok/s
+  [eval step 1000] val_loss 2.3405 (best 2.3405)
+  [gen] Once upon a time,  in a small house, there was a very special place to a big tree near a field of the sea. The boat had many friends there, like animals and trees. They all wanted to a place to go on a journey with the other animals to play on the big big farm.
+  [gen] The little girl wanted to  help the girl. The cat asked the mouse, "Can I help you find its friend."
+  [gen] In the dark forest,  there was a little bird with the bird in the tree's nest near a big garden. The big tree looked sad and could not find anything to eat.
+  saved ckpt/tiny25m/model_1000.pt
+step 1025/5000 loss 2.2901 lr 9.60e-05 260 tok/s
+step 1050/5000 loss 2.1281 lr 9.56e-05 427 tok/s
+step 1075/5000 loss 2.3848 lr 9.53e-05 424 tok/s
+step 1100/5000 loss 2.0510 lr 9.50e-05 427 tok/s
+step 1125/5000 loss 2.5843 lr 9.46e-05 426 tok/s
+step 1150/5000 loss 2.5651 lr 9.42e-05 422 tok/s
+step 1175/5000 loss 2.5713 lr 9.38e-05 421 tok/s
+step 1200/5000 loss 2.4041 lr 9.34e-05 422 tok/s
+step 1225/5000 loss 2.2352 lr 9.30e-05 427 tok/s
+step 1250/5000 loss 2.2606 lr 9.26e-05 423 tok/s
+step 1275/5000 loss 2.2703 lr 9.22e-05 422 tok/s
+step 1300/5000 loss 2.4597 lr 9.18e-05 422 tok/s
+step 1325/5000 loss 2.4358 lr 9.13e-05 422 tok/s
+step 1350/5000 loss 2.3806 lr 9.09e-05 426 tok/s
+step 1375/5000 loss 2.3900 lr 9.04e-05 420 tok/s
+step 1400/5000 loss 2.3329 lr 8.99e-05 422 tok/s
+step 1425/5000 loss 2.2474 lr 8.94e-05 422 tok/s
+step 1450/5000 loss 2.3502 lr 8.89e-05 421 tok/s
+step 1475/5000 loss 2.4066 lr 8.84e-05 422 tok/s
+step 1500/5000 loss 2.3801 lr 8.79e-05 420 tok/s
+  [eval step 1500] val_loss 2.3284 (best 2.3284)
+  [gen] Once upon a time,  there lived a little boy named Tim. Every day, he would play with his friends and help the kids in the town.
+  [gen] The little girl wanted to  help the tall man. The pretty mouse wanted to help.
+  [gen] In the dark forest,  all the animals in the big sea, and their friends. The other animals in the trees in the town, the fish, and all their houses, green things, more flowers, and people in the forest.
+step 1525/5000 loss 2.3024 lr 8.74e-05 265 tok/s
+step 1550/5000 loss 2.3755 lr 8.68e-05 422 tok/s
+step 1575/5000 loss 2.3612 lr 8.63e-05 418 tok/s
+step 1600/5000 loss 2.1853 lr 8.57e-05 419 tok/s
+step 1625/5000 loss 2.3783 lr 8.51e-05 421 tok/s
+step 1650/5000 loss 2.5388 lr 8.46e-05 419 tok/s
+step 1675/5000 loss 1.8248 lr 8.40e-05 420 tok/s
+step 1700/5000 loss 2.4070 lr 8.34e-05 421 tok/s
+step 1725/5000 loss 2.0517 lr 8.28e-05 421 tok/s
+step 1750/5000 loss 2.4317 lr 8.22e-05 421 tok/s
+step 1775/5000 loss 2.4273 lr 8.16e-05 422 tok/s
+step 1800/5000 loss 2.2379 lr 8.10e-05 421 tok/s
+step 1825/5000 loss 2.2501 lr 8.03e-05 421 tok/s
+step 1850/5000 loss 2.2434 lr 7.97e-05 421 tok/s
+step 1875/5000 loss 1.9137 lr 7.90e-05 422 tok/s
+step 1900/5000 loss 2.2223 lr 7.84e-05 389 tok/s
+step 1925/5000 loss 2.2548 lr 7.77e-05 389 tok/s
+step 1950/5000 loss 2.4698 lr 7.71e-05 388 tok/s
+step 1975/5000 loss 1.9972 lr 7.64e-05 382 tok/s
+step 2000/5000 loss 2.4020 lr 7.57e-05 376 tok/s
+  [eval step 2000] val_loss 2.3080 (best 2.3080)
+  [gen] Once upon a time,  in a big green park and lived in a big forest. Every day, the bird would sing all by his fish all around the pretty flowers. One day, he went to the tree to find a new there. He decided to open the door and it was not so colorful!
+  [gen] The little girl wanted to  help the little mouse. The small cat said, "I want to buy a yummy cake! I will use your food and toys." The bird told the cat about the old man, the girl with the cat's sack.
+  [gen] In the dark forest,  there was an elderly owl who lived in a small, brown bear. He loved to walk and play with his friends, from the river, all around the town always kept walking in the woods near their house.
+  saved ckpt/tiny25m/model_2000.pt
+step 2025/5000 loss 2.3502 lr 7.50e-05 229 tok/s
+step 2050/5000 loss 2.3404 lr 7.43e-05 391 tok/s
+step 2075/5000 loss 2.1333 lr 7.36e-05 386 tok/s
+step 2100/5000 loss 2.4253 lr 7.29e-05 391 tok/s
+step 2125/5000 loss 2.0974 lr 7.22e-05 391 tok/s
+step 2150/5000 loss 2.1223 lr 7.15e-05 403 tok/s
+step 2175/5000 loss 2.2124 lr 7.08e-05 436 tok/s
+step 2200/5000 loss 2.4553 lr 7.01e-05 436 tok/s
+step 2225/5000 loss 2.1679 lr 6.93e-05 438 tok/s
+step 2250/5000 loss 2.0824 lr 6.86e-05 439 tok/s
+step 2275/5000 loss 2.2627 lr 6.79e-05 429 tok/s
+step 2300/5000 loss 2.7777 lr 6.71e-05 436 tok/s
+step 2325/5000 loss 2.2637 lr 6.64e-05 421 tok/s
+step 2350/5000 loss 2.1451 lr 6.57e-05 437 tok/s

logs/train_phase2b_g25_final.log

@@ -0,0 +1,28 @@
+resumed ckpt/tiny25m/model_4500.pt at step 4500
+params: 25.43M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 4525/5000 loss 2.1578 lr 1.23e-05 300 tok/s
+step 4550/5000 loss 2.1618 lr 1.21e-05 303 tok/s
+step 4575/5000 loss 2.0825 lr 1.19e-05 345 tok/s
+step 4600/5000 loss 2.3762 lr 1.17e-05 327 tok/s
+step 4625/5000 loss 2.0607 lr 1.15e-05 326 tok/s
+step 4650/5000 loss 1.9639 lr 1.13e-05 307 tok/s
+step 4675/5000 loss 2.0407 lr 1.11e-05 319 tok/s
+step 4700/5000 loss 2.1899 lr 1.09e-05 321 tok/s
+step 4725/5000 loss 2.6930 lr 1.08e-05 315 tok/s
+step 4750/5000 loss 2.1291 lr 1.07e-05 331 tok/s
+step 4775/5000 loss 2.4192 lr 1.05e-05 339 tok/s
+step 4800/5000 loss 2.2638 lr 1.04e-05 375 tok/s
+step 4825/5000 loss 2.1685 lr 1.03e-05 427 tok/s
+step 4850/5000 loss 2.2833 lr 1.02e-05 412 tok/s
+step 4875/5000 loss 2.2088 lr 1.02e-05 425 tok/s
+step 4900/5000 loss 2.1150 lr 1.01e-05 343 tok/s
+step 4925/5000 loss 2.3808 lr 1.01e-05 335 tok/s
+step 4950/5000 loss 2.2054 lr 1.00e-05 356 tok/s
+step 4975/5000 loss 2.0439 lr 1.00e-05 387 tok/s
+step 5000/5000 loss 2.1454 lr 1.00e-05 400 tok/s
+  [eval step 5000] val_loss 2.2294 (best 2.2294)
+  [gen] Once upon a time,  in a big tree near the tree to the tall tree and helped many things in the water.
+  [gen] The little girl wanted to  help the little boy. The kind man wanted to help a nice lady. The old man gave the kind boy some food from the man. The man was happy to have the man's help. He went home and told the kind man about his work. The man helped to fix things with the fire truck.
+  [gen] In the dark forest,  there was a big tree with lots of plants and trees. All the animals in the forest were happy to see the nest near their home like the sun, and the plants in the sun and water. In the end, all of the animals saw the flowers, they found in the room, and the leaves that covered on the ground.
+  saved ckpt/tiny25m/model_5000.pt
+done

logs/train_phase2b_r2.log

@@ -0,0 +1,68 @@
+resumed ckpt/tiny18m2/model_1500.pt at step 1500
+params: 16.77M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 1525/8500 loss 2.6791 lr 9.45e-05 754 tok/s
+step 1550/8500 loss 2.5554 lr 9.43e-05 695 tok/s
+step 1575/8500 loss 2.6010 lr 9.40e-05 689 tok/s
+step 1600/8500 loss 2.4354 lr 9.38e-05 684 tok/s
+step 1625/8500 loss 2.7150 lr 9.36e-05 685 tok/s
+step 1650/8500 loss 2.8348 lr 9.34e-05 682 tok/s
+step 1675/8500 loss 2.1119 lr 9.32e-05 684 tok/s
+step 1700/8500 loss 2.6179 lr 9.29e-05 684 tok/s
+step 1725/8500 loss 2.2416 lr 9.27e-05 686 tok/s
+step 1750/8500 loss 2.7505 lr 9.25e-05 683 tok/s
+step 1775/8500 loss 2.6319 lr 9.22e-05 688 tok/s
+step 1800/8500 loss 2.4757 lr 9.20e-05 716 tok/s
+step 1825/8500 loss 2.5563 lr 9.18e-05 710 tok/s
+step 1850/8500 loss 2.5105 lr 9.15e-05 668 tok/s
+step 1875/8500 loss 2.0956 lr 9.13e-05 713 tok/s
+step 1900/8500 loss 2.4147 lr 9.10e-05 716 tok/s
+step 1925/8500 loss 2.4454 lr 9.07e-05 713 tok/s
+step 1950/8500 loss 2.7386 lr 9.05e-05 713 tok/s
+step 1975/8500 loss 2.2794 lr 9.02e-05 713 tok/s
+step 2000/8500 loss 2.6433 lr 9.00e-05 678 tok/s
+  [eval step 2000] val_loss 2.4536 (best 2.4536)
+  [gen] Once upon a time,  in a small house, a little girl named Sue. They lived in a small town with the animals under the tree. They were happy to play together and have fun with their toys.
+  [gen] The little girl wanted to  help the lady. The boy said, "My name is an old chair. I can't find it."
+  [gen] In the dark forest,  there was a little bird who wanted to fly near the tree. So, they went out of the race and saw a small bird on the top. The bird thought up with an idea. The stick said, "I can't know how to get my home."
+  saved ckpt/tiny18m2/model_2000.pt
+step 2025/8500 loss 2.6009 lr 8.97e-05 433 tok/s
+step 2050/8500 loss 2.5285 lr 8.94e-05 681 tok/s
+step 2075/8500 loss 2.3518 lr 8.91e-05 692 tok/s
+step 2100/8500 loss 2.6506 lr 8.89e-05 703 tok/s
+step 2125/8500 loss 2.3366 lr 8.86e-05 718 tok/s
+step 2150/8500 loss 2.3934 lr 8.83e-05 725 tok/s
+step 2175/8500 loss 2.5290 lr 8.80e-05 721 tok/s
+step 2200/8500 loss 2.6982 lr 8.77e-05 721 tok/s
+step 2225/8500 loss 2.3802 lr 8.74e-05 722 tok/s
+step 2250/8500 loss 2.2969 lr 8.71e-05 723 tok/s
+step 2275/8500 loss 2.4702 lr 8.68e-05 724 tok/s
+step 2300/8500 loss 3.1186 lr 8.65e-05 715 tok/s
+step 2325/8500 loss 2.4230 lr 8.62e-05 720 tok/s
+step 2350/8500 loss 2.2874 lr 8.59e-05 712 tok/s
+step 2375/8500 loss 2.4387 lr 8.56e-05 632 tok/s
+step 2400/8500 loss 2.2299 lr 8.53e-05 739 tok/s
+step 2425/8500 loss 2.4471 lr 8.50e-05 693 tok/s
+step 2450/8500 loss 2.3456 lr 8.46e-05 695 tok/s
+step 2475/8500 loss 2.3106 lr 8.43e-05 730 tok/s
+step 2500/8500 loss 2.3341 lr 8.40e-05 714 tok/s
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lm.py", line 248, in <module>
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lm.py", line 231, in main
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train/train_lm.py", line 167, in save
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/serialization.py", line 1002, in save
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/serialization.py", line 864, in _open_zipfile_writer
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/serialization.py", line 828, in __init__
+RuntimeError: Parent directory ckpt/tiny18m2 does not exist.
+Exception ignored in atexit callback: <function dump_compile_times at 0x7c156e7f60>
+Traceback (most recent call last):
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/_dynamo/utils.py", line 925, in dump_compile_times
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/_dynamo/utils.py", line 911, in compile_times
+  File "/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/.venv/lib/python3.12/site-packages/torch/_dynamo/utils.py", line 249, in tabulate
+  File "<frozen importlib._bootstrap>", line 1360, in _find_and_load
+  File "<frozen importlib._bootstrap>", line 1322, in _find_and_load_unlocked
+  File "<frozen importlib._bootstrap>", line 1262, in _find_spec
+  File "<frozen importlib._bootstrap_external>", line 1528, in find_spec
+  File "<frozen importlib._bootstrap_external>", line 1502, in _get_spec
+  File "<frozen importlib._bootstrap_external>", line 1605, in find_spec
+  File "<frozen importlib._bootstrap_external>", line 1648, in _fill_cache
+OSError: [Errno 38] Function not implemented: '/root/Documents/Codex/2026-07-31/so-i-ve-got-a-task/train'

logs/train_phase2b_r3.log

@@ -0,0 +1,321 @@
+resumed ckpt/tiny18m2/model_2000.pt at step 2000
+params: 16.77M  train_tokens: 32,562,432  steps_per_epoch: 7949
+step 2025/8500 loss 2.6009 lr 8.97e-05 768 tok/s
+step 2050/8500 loss 2.5285 lr 8.94e-05 787 tok/s
+step 2075/8500 loss 2.3518 lr 8.91e-05 777 tok/s
+step 2100/8500 loss 2.6506 lr 8.89e-05 714 tok/s
+step 2125/8500 loss 2.3366 lr 8.86e-05 710 tok/s
+step 2150/8500 loss 2.3934 lr 8.83e-05 683 tok/s
+step 2175/8500 loss 2.5290 lr 8.80e-05 712 tok/s
+step 2200/8500 loss 2.6982 lr 8.77e-05 721 tok/s
+step 2225/8500 loss 2.3802 lr 8.74e-05 721 tok/s
+step 2250/8500 loss 2.2969 lr 8.71e-05 733 tok/s
+step 2275/8500 loss 2.4702 lr 8.68e-05 743 tok/s
+step 2300/8500 loss 3.1186 lr 8.65e-05 718 tok/s
+step 2325/8500 loss 2.4230 lr 8.62e-05 749 tok/s
+step 2350/8500 loss 2.2874 lr 8.59e-05 670 tok/s
+step 2375/8500 loss 2.4387 lr 8.56e-05 682 tok/s
+step 2400/8500 loss 2.2299 lr 8.53e-05 740 tok/s
+step 2425/8500 loss 2.4471 lr 8.50e-05 734 tok/s
+step 2450/8500 loss 2.3456 lr 8.46e-05 734 tok/s
+step 2475/8500 loss 2.3106 lr 8.43e-05 735 tok/s
+step 2500/8500 loss 2.3341 lr 8.40e-05 736 tok/s
+  [eval step 2500] val_loss 2.4288 (best 2.4288)
+  [gen] Once upon a time,  there was a big green tree with many animals. The leaf lived in a small house with the pond under the trees and bugs. They liked to play near the big tree. Every day, the bird would sit there for water on their tree to go fast, like the other fish that they went for it.
+  [gen] The little girl wanted to  help the little bird. The big bird saw the cat and wanted to play with its family too.
+  [gen] In the dark forest,  there was a little bird who had found an old rock in the tree. The tree thought it could be like the right thing and wanted to help the bird find its home.
+step 2525/8500 loss 2.3900 lr 8.37e-05 457 tok/s
+step 2550/8500 loss 2.3513 lr 8.33e-05 643 tok/s
+step 2575/8500 loss 2.4871 lr 8.30e-05 724 tok/s
+step 2600/8500 loss 2.6928 lr 8.27e-05 730 tok/s
+step 2625/8500 loss 2.2624 lr 8.23e-05 728 tok/s
+step 2650/8500 loss 2.4440 lr 8.20e-05 729 tok/s
+step 2675/8500 loss 2.5323 lr 8.17e-05 730 tok/s
+step 2700/8500 loss 2.5437 lr 8.13e-05 724 tok/s
+step 2725/8500 loss 2.3197 lr 8.10e-05 722 tok/s
+step 2750/8500 loss 2.7255 lr 8.06e-05 720 tok/s
+step 2775/8500 loss 2.4288 lr 8.03e-05 711 tok/s
+step 2800/8500 loss 2.5065 lr 7.99e-05 737 tok/s
+step 2825/8500 loss 2.2189 lr 7.96e-05 795 tok/s
+step 2850/8500 loss 2.4310 lr 7.92e-05 733 tok/s
+step 2875/8500 loss 2.1453 lr 7.88e-05 739 tok/s
+step 2900/8500 loss 2.2286 lr 7.85e-05 685 tok/s
+step 2925/8500 loss 2.5208 lr 7.81e-05 729 tok/s
+step 2950/8500 loss 2.6875 lr 7.77e-05 681 tok/s
+step 2975/8500 loss 2.5541 lr 7.74e-05 708 tok/s
+step 3000/8500 loss 2.5958 lr 7.70e-05 736 tok/s
+  [eval step 3000] val_loss 2.4096 (best 2.4096)
+  [gen] Once upon a time,  in a big green park. It was in the sky. In this tree, there lived an old bear named Mark and his family. They were all all together and they played.
+  [gen] The little girl wanted to  help the fox. The man told the dog about the big cat. The squirrel wanted to help Tim find its home in the tree.
+  [gen] In the dark forest,  there was a little bird who lived in a pond. The bird loved to play with his friends and they all played together.
+  saved ckpt/tiny18m2/model_3000.pt
+step 3025/8500 loss 2.3156 lr 7.66e-05 437 tok/s
+step 3050/8500 loss 2.4188 lr 7.63e-05 737 tok/s
+step 3075/8500 loss 2.3696 lr 7.59e-05 724 tok/s
+step 3100/8500 loss 2.6240 lr 7.55e-05 709 tok/s
+step 3125/8500 loss 2.6688 lr 7.51e-05 708 tok/s
+step 3150/8500 loss 2.2894 lr 7.47e-05 713 tok/s
+step 3175/8500 loss 2.4838 lr 7.44e-05 725 tok/s
+step 3200/8500 loss 2.1285 lr 7.40e-05 725 tok/s
+step 3225/8500 loss 2.3186 lr 7.36e-05 725 tok/s
+step 3250/8500 loss 2.3504 lr 7.32e-05 722 tok/s
+step 3275/8500 loss 2.2351 lr 7.28e-05 718 tok/s
+step 3300/8500 loss 2.5447 lr 7.24e-05 672 tok/s
+step 3325/8500 loss 2.5077 lr 7.20e-05 688 tok/s
+step 3350/8500 loss 2.0609 lr 7.16e-05 701 tok/s
+step 3375/8500 loss 2.4588 lr 7.12e-05 700 tok/s
+step 3400/8500 loss 2.3440 lr 7.08e-05 656 tok/s
+step 3425/8500 loss 2.3952 lr 7.04e-05 675 tok/s
+step 3450/8500 loss 2.5002 lr 7.00e-05 719 tok/s
+step 3475/8500 loss 2.4306 lr 6.96e-05 725 tok/s
+step 3500/8500 loss 2.6222 lr 6.92e-05 731 tok/s
+  [eval step 3500] val_loss 2.3932 (best 2.3932)
+  [gen] Once upon a time,  there was a little bird who was sad because he lost his home.
+  [gen] The little girl wanted to  help the cat. The bird saw a big tree with many branches. The bird had no leaves and it was stuck in the box. The big dog helped the mouse reach up and picked them up with his beak.
+  [gen] In the dark forest,  there was a boy with a big bag with lots of toys and games. Many other kids would look for different shapes and shapes on them to people. All the animals were happy in the village because they only wanted to see new things.
+step 3525/8500 loss 2.1995 lr 6.88e-05 459 tok/s
+step 3550/8500 loss 2.4128 lr 6.84e-05 721 tok/s
+step 3575/8500 loss 2.1446 lr 6.80e-05 730 tok/s
+step 3600/8500 loss 2.2230 lr 6.76e-05 742 tok/s
+step 3625/8500 loss 2.4896 lr 6.72e-05 807 tok/s
+step 3650/8500 loss 2.2890 lr 6.68e-05 814 tok/s
+step 3675/8500 loss 2.4059 lr 6.64e-05 812 tok/s
+step 3700/8500 loss 2.4369 lr 6.60e-05 810 tok/s
+step 3725/8500 loss 2.6373 lr 6.55e-05 815 tok/s
+step 3750/8500 loss 2.2640 lr 6.51e-05 782 tok/s
+step 3775/8500 loss 2.3666 lr 6.47e-05 729 tok/s
+step 3800/8500 loss 2.2370 lr 6.43e-05 749 tok/s
+step 3825/8500 loss 2.1752 lr 6.39e-05 732 tok/s
+step 3850/8500 loss 2.4906 lr 6.35e-05 727 tok/s
+step 3875/8500 loss 2.3461 lr 6.30e-05 716 tok/s
+step 3900/8500 loss 2.4888 lr 6.26e-05 602 tok/s
+step 3925/8500 loss 2.2797 lr 6.22e-05 661 tok/s
+step 3950/8500 loss 2.5380 lr 6.18e-05 612 tok/s
+step 3975/8500 loss 2.5046 lr 6.14e-05 761 tok/s
+step 4000/8500 loss 2.2241 lr 6.09e-05 669 tok/s
+  [eval step 4000] val_loss 2.3799 (best 2.3799)
+  [gen] Once upon a time,  in a big blue forest lived in the big tree in the water. They all liked to play and have fun together.
+  [gen] The little girl wanted to  help the little cat. The big bird told the big cat, a small dog, for help, but the big dog, the cat, the small cat, the dog, and the squirrel were together.
+  [gen] In the dark forest,  there was a kind old owl named Sam, came up to help Tim, and they walked together to find food to find something. They decided to buy some food from their friends, the big, wise wise old man, who could not tell him about his new house, and they all took them home with their little brother, the two of them loved his new toy back and gave her a big,
+  saved ckpt/tiny18m2/model_4000.pt
+step 4025/8500 loss 2.4218 lr 6.05e-05 427 tok/s
+step 4050/8500 loss 2.3075 lr 6.01e-05 693 tok/s
+step 4075/8500 loss 2.4060 lr 5.97e-05 692 tok/s
+step 4100/8500 loss 2.3822 lr 5.93e-05 690 tok/s
+step 4125/8500 loss 2.3735 lr 5.88e-05 699 tok/s
+step 4150/8500 loss 2.5614 lr 5.84e-05 716 tok/s
+step 4175/8500 loss 2.1389 lr 5.80e-05 713 tok/s
+step 4200/8500 loss 2.3110 lr 5.76e-05 724 tok/s
+step 4225/8500 loss 2.2894 lr 5.71e-05 726 tok/s
+step 4250/8500 loss 2.5321 lr 5.67e-05 666 tok/s
+step 4275/8500 loss 2.5922 lr 5.63e-05 704 tok/s
+step 4300/8500 loss 2.1300 lr 5.59e-05 698 tok/s
+step 4325/8500 loss 2.5078 lr 5.54e-05 697 tok/s
+step 4350/8500 loss 2.4255 lr 5.50e-05 698 tok/s
+step 4375/8500 loss 2.3638 lr 5.46e-05 696 tok/s
+step 4400/8500 loss 2.2405 lr 5.41e-05 695 tok/s
+step 4425/8500 loss 2.3808 lr 5.37e-05 695 tok/s
+step 4450/8500 loss 2.2644 lr 5.33e-05 693 tok/s
+step 4475/8500 loss 2.5325 lr 5.29e-05 696 tok/s
+step 4500/8500 loss 2.3401 lr 5.24e-05 694 tok/s
+  [eval step 4500] val_loss 2.3620 (best 2.3620)
+  [gen] Once upon a time,  there lived a little bird. The fly around in the big forest, and they were all very happy to have lots of fun together. They always played together every day, and they would come out for more adventures together. When they were done, they started to play and laugh with them all day long.
+  [gen] The little girl wanted to  help the little bird. The wise owl said, "I will help you find food."
+  [gen] In the dark forest,  there was a little girl named Lucy. She lived in a small house with a pretty home near their house. The farm had many friends who liked to play and have fun together.
+step 4525/8500 loss 2.4057 lr 5.20e-05 426 tok/s
+step 4550/8500 loss 2.3704 lr 5.16e-05 716 tok/s
+step 4575/8500 loss 2.2685 lr 5.12e-05 740 tok/s
+step 4600/8500 loss 2.6391 lr 5.07e-05 741 tok/s
+step 4625/8500 loss 2.2122 lr 5.03e-05 738 tok/s
+step 4650/8500 loss 2.1762 lr 4.99e-05 743 tok/s
+step 4675/8500 loss 2.2390 lr 4.95e-05 738 tok/s
+step 4700/8500 loss 2.3194 lr 4.91e-05 738 tok/s
+step 4725/8500 loss 2.9141 lr 4.86e-05 740 tok/s
+step 4750/8500 loss 2.2902 lr 4.82e-05 739 tok/s
+step 4775/8500 loss 2.5889 lr 4.78e-05 744 tok/s
+step 4800/8500 loss 2.4242 lr 4.74e-05 740 tok/s
+step 4825/8500 loss 2.3526 lr 4.70e-05 740 tok/s
+step 4850/8500 loss 2.4663 lr 4.65e-05 743 tok/s
+step 4875/8500 loss 2.4006 lr 4.61e-05 742 tok/s
+step 4900/8500 loss 2.2832 lr 4.57e-05 742 tok/s
+step 4925/8500 loss 2.5359 lr 4.53e-05 741 tok/s
+step 4950/8500 loss 2.3870 lr 4.49e-05 744 tok/s
+step 4975/8500 loss 2.1696 lr 4.45e-05 740 tok/s
+step 5000/8500 loss 2.3123 lr 4.40e-05 741 tok/s
+  [eval step 5000] val_loss 2.3509 (best 2.3509)
+  [gen] Once upon a time,  in a small house, there lived a big, big tree in the sea. The big purple boat was very big and soft, and the blue green grass were red, and yellow, and green clouds. The little blue bird loved to play with his friends in the park every day, and they would zoom around with its soft fur, yellow and shiny rocks.
+  [gen] The little girl wanted to  help the little bird.
+  [gen] In the dark forest,  there was a little boy named Tim. Tim lived in a town with many people and they all wanted to have fun with them.
+  saved ckpt/tiny18m2/model_5000.pt
+step 5025/8500 loss 2.4307 lr 4.36e-05 473 tok/s
+step 5050/8500 loss 2.1796 lr 4.32e-05 741 tok/s
+step 5075/8500 loss 2.2374 lr 4.28e-05 741 tok/s
+step 5100/8500 loss 2.2539 lr 4.24e-05 738 tok/s
+step 5125/8500 loss 2.3968 lr 4.20e-05 745 tok/s
+step 5150/8500 loss 2.1371 lr 4.16e-05 748 tok/s
+step 5175/8500 loss 2.3619 lr 4.12e-05 742 tok/s
+step 5200/8500 loss 2.3210 lr 4.08e-05 740 tok/s
+step 5225/8500 loss 2.6705 lr 4.04e-05 744 tok/s
+step 5250/8500 loss 2.2133 lr 4.00e-05 739 tok/s
+step 5275/8500 loss 2.3599 lr 3.96e-05 740 tok/s
+step 5300/8500 loss 2.3256 lr 3.92e-05 746 tok/s
+step 5325/8500 loss 2.6780 lr 3.88e-05 742 tok/s
+step 5350/8500 loss 2.6026 lr 3.84e-05 744 tok/s
+step 5375/8500 loss 2.0700 lr 3.80e-05 748 tok/s
+step 5400/8500 loss 2.4377 lr 3.76e-05 743 tok/s
+step 5425/8500 loss 2.5846 lr 3.72e-05 741 tok/s
+step 5450/8500 loss 1.9803 lr 3.68e-05 742 tok/s
+step 5475/8500 loss 2.1106 lr 3.64e-05 743 tok/s
+step 5500/8500 loss 2.2869 lr 3.60e-05 738 tok/s
+  [eval step 5500] val_loss 2.3343 (best 2.3343)
+  [gen] Once upon a time,  in a small house, there lived a boy named Tim. He loved to play with his friends at the park. One day, while playing, he saw an even bigger fire in the sky!
+  [gen] The little girl wanted to  help the little girl. The big cat told the cat about the little dog.
+  [gen] In the dark forest,  there was a tiny bird. The bird could not fly high in this tree. It had many leaves that were sad. The bird wanted to be like its friends and they all his friends loved.
+step 5525/8500 loss 2.5666 lr 3.56e-05 477 tok/s
+step 5550/8500 loss 2.6961 lr 3.53e-05 745 tok/s
+step 5575/8500 loss 2.2242 lr 3.49e-05 739 tok/s
+step 5600/8500 loss 2.5946 lr 3.45e-05 741 tok/s
+step 5625/8500 loss 2.4833 lr 3.41e-05 741 tok/s
+step 5650/8500 loss 2.3140 lr 3.37e-05 740 tok/s
+step 5675/8500 loss 2.4133 lr 3.34e-05 743 tok/s
+step 5700/8500 loss 2.0494 lr 3.30e-05 704 tok/s
+step 5725/8500 loss 2.4197 lr 3.26e-05 750 tok/s
+step 5750/8500 loss 2.5361 lr 3.23e-05 753 tok/s
+step 5775/8500 loss 1.8925 lr 3.19e-05 746 tok/s
+step 5800/8500 loss 2.3718 lr 3.15e-05 729 tok/s
+step 5825/8500 loss 2.4211 lr 3.12e-05 717 tok/s
+step 5850/8500 loss 2.1941 lr 3.08e-05 672 tok/s
+step 5875/8500 loss 2.7147 lr 3.04e-05 676 tok/s
+step 5900/8500 loss 2.3902 lr 3.01e-05 689 tok/s
+step 5925/8500 loss 2.1891 lr 2.97e-05 731 tok/s
+step 5950/8500 loss 2.2179 lr 2.94e-05 696 tok/s
+step 5975/8500 loss 2.4671 lr 2.90e-05 677 tok/s
+step 6000/8500 loss 2.2648 lr 2.87e-05 677 tok/s
+  [eval step 6000] val_loss 2.3256 (best 2.3256)
+  [gen] Once upon a time,  a little boy had a new friend named Tim. They played all day and had lots of fun together in the big, soft lake.
+  [gen] The little girl wanted to  help the little girl. The small cat saw a tiny bug. The little boy said, "Hi, pretty bird! I will teach you about my food. Why are you sad?"
+  [gen] In the dark forest,  there was a little bird sitting on a tree. The big bird lived in the nest near the small nest. The bird wanted to help. They flew together in the sky and the sun was bright in the wind.
+  saved ckpt/tiny18m2/model_6000.pt
+step 6025/8500 loss 2.2873 lr 2.83e-05 439 tok/s
+step 6050/8500 loss 2.3827 lr 2.80e-05 708 tok/s
+step 6075/8500 loss 2.2150 lr 2.77e-05 708 tok/s
+step 6100/8500 loss 2.0553 lr 2.73e-05 706 tok/s
+step 6125/8500 loss 2.1190 lr 2.70e-05 707 tok/s
+step 6150/8500 loss 2.4093 lr 2.67e-05 708 tok/s
+step 6175/8500 loss 2.0879 lr 2.63e-05 705 tok/s
+step 6200/8500 loss 2.1990 lr 2.60e-05 709 tok/s
+step 6225/8500 loss 2.4327 lr 2.57e-05 705 tok/s
+step 6250/8500 loss 2.5909 lr 2.54e-05 667 tok/s
+step 6275/8500 loss 1.9986 lr 2.50e-05 709 tok/s
+step 6300/8500 loss 2.2856 lr 2.47e-05 746 tok/s
+step 6325/8500 loss 2.8116 lr 2.44e-05 745 tok/s
+step 6350/8500 loss 2.5897 lr 2.41e-05 749 tok/s
+step 6375/8500 loss 2.3566 lr 2.38e-05 745 tok/s
+step 6400/8500 loss 2.2315 lr 2.35e-05 723 tok/s
+step 6425/8500 loss 2.2787 lr 2.32e-05 675 tok/s
+step 6450/8500 loss 2.3502 lr 2.29e-05 572 tok/s
+step 6475/8500 loss 2.2950 lr 2.26e-05 572 tok/s
+step 6500/8500 loss 2.1582 lr 2.23e-05 592 tok/s
+  [eval step 6500] val_loss 2.3158 (best 2.3158)
+  [gen] Once upon a time,  there lived a little fish named Fin. Spot loved to swim with his big green pond in the sea. The deep nest was warm and strong. The bird would fly high up high and down on high tree. All the animals in the water liked to play near the big tree.
+  [gen] The little girl wanted to  help the cat. The mouse saw the cat and wanted to help.
+  [gen] In the dark forest,  there was a little fish who lived near the big boat. They all liked to join in the fish and they all wanted to play together.
+step 6525/8500 loss 2.3357 lr 2.20e-05 382 tok/s
+step 6550/8500 loss 2.5360 lr 2.17e-05 604 tok/s
+step 6575/8500 loss 2.1777 lr 2.14e-05 652 tok/s
+step 6600/8500 loss 2.3372 lr 2.11e-05 709 tok/s
+step 6625/8500 loss 2.2236 lr 2.09e-05 765 tok/s
+step 6650/8500 loss 2.2979 lr 2.06e-05 753 tok/s
+step 6675/8500 loss 1.9714 lr 2.03e-05 747 tok/s
+step 6700/8500 loss 2.5266 lr 2.00e-05 747 tok/s
+step 6725/8500 loss 2.4799 lr 1.98e-05 749 tok/s
+step 6750/8500 loss 2.1943 lr 1.95e-05 752 tok/s
+step 6775/8500 loss 2.6434 lr 1.93e-05 740 tok/s
+step 6800/8500 loss 2.0762 lr 1.90e-05 750 tok/s
+step 6825/8500 loss 2.5136 lr 1.87e-05 749 tok/s
+step 6850/8500 loss 2.2378 lr 1.85e-05 747 tok/s
+step 6875/8500 loss 2.3700 lr 1.82e-05 746 tok/s
+step 6900/8500 loss 2.4945 lr 1.80e-05 747 tok/s
+step 6925/8500 loss 2.3929 lr 1.78e-05 393 tok/s
+step 6950/8500 loss 2.1891 lr 1.75e-05 224 tok/s
+step 6975/8500 loss 2.7453 lr 1.73e-05 411 tok/s
+step 7000/8500 loss 2.1759 lr 1.71e-05 240 tok/s
+  [eval step 7000] val_loss 2.3082 (best 2.3082)
+  [gen] Once upon a time,  there was a big tree in the sky. The tree liked to sing and play all day long. One sunny day, it started to rain. It was warm and wind blew on the cloud away. It felt so strong and strong.
+  [gen] The little girl wanted to  help the little cat. The small bird flew down and asked the kind old man, "Don't be scared of a special treat! I will help your friends."
+  [gen] In the dark forest,  there was a kind and beautiful sun. The sun was shining, it was bright, and it made everyone happy.
+  saved ckpt/tiny18m2/model_7000.pt
+step 7025/8500 loss 2.2284 lr 1.68e-05 441 tok/s
+step 7050/8500 loss 2.4942 lr 1.66e-05 722 tok/s
+step 7075/8500 loss 2.3786 lr 1.64e-05 716 tok/s
+step 7100/8500 loss 2.2349 lr 1.62e-05 718 tok/s
+step 7125/8500 loss 2.3990 lr 1.60e-05 712 tok/s
+step 7150/8500 loss 2.2449 lr 1.57e-05 699 tok/s
+step 7175/8500 loss 2.3882 lr 1.55e-05 645 tok/s
+step 7200/8500 loss 2.4181 lr 1.53e-05 744 tok/s
+step 7225/8500 loss 2.5697 lr 1.51e-05 774 tok/s
+step 7250/8500 loss 2.0296 lr 1.49e-05 780 tok/s
+step 7275/8500 loss 2.4014 lr 1.48e-05 768 tok/s
+step 7300/8500 loss 1.9945 lr 1.46e-05 761 tok/s
+step 7325/8500 loss 2.5492 lr 1.44e-05 759 tok/s
+step 7350/8500 loss 2.2305 lr 1.42e-05 776 tok/s
+step 7375/8500 loss 2.3821 lr 1.40e-05 775 tok/s
+step 7400/8500 loss 2.3064 lr 1.38e-05 776 tok/s
+step 7425/8500 loss 2.5148 lr 1.37e-05 779 tok/s
+step 7450/8500 loss 2.0114 lr 1.35e-05 777 tok/s
+step 7475/8500 loss 2.3851 lr 1.33e-05 864 tok/s
+step 7500/8500 loss 2.5159 lr 1.32e-05 768 tok/s
+  [eval step 7500] val_loss 2.3027 (best 2.3027)
+  [gen] Once upon a time,  in a big, red field of trees, there was a beautiful picture full of tall trees, and even yellow flowers.
+  [gen] The little girl wanted to  see the pretty bug. The bird thought the bird could fly in the tree. The little bird wanted to help the bird.
+  [gen] In the dark forest,  there was a little bunny named Bobby. Bobo was a new friend, but she had an idea to help his new friend, a big bird and they all took turns dancing together.
+step 7525/8500 loss 2.4616 lr 1.30e-05 483 tok/s
+step 7550/8500 loss 2.3869 lr 1.29e-05 737 tok/s
+step 7575/8500 loss 2.3460 lr 1.27e-05 739 tok/s
+step 7600/8500 loss 2.2501 lr 1.26e-05 741 tok/s
+step 7625/8500 loss 2.4396 lr 1.24e-05 743 tok/s
+step 7650/8500 loss 2.1035 lr 1.23e-05 687 tok/s
+step 7675/8500 loss 2.2924 lr 1.22e-05 677 tok/s
+step 7700/8500 loss 2.2460 lr 1.20e-05 671 tok/s
+step 7725/8500 loss 2.3714 lr 1.19e-05 782 tok/s
+step 7750/8500 loss 2.1824 lr 1.18e-05 747 tok/s
+step 7775/8500 loss 2.3385 lr 1.17e-05 729 tok/s
+step 7800/8500 loss 2.3321 lr 1.16e-05 738 tok/s
+step 7825/8500 loss 2.1710 lr 1.15e-05 676 tok/s
+step 7850/8500 loss 2.3334 lr 1.14e-05 724 tok/s
+step 7875/8500 loss 2.1751 lr 1.13e-05 738 tok/s
+step 7900/8500 loss 2.4071 lr 1.12e-05 735 tok/s
+step 7925/8500 loss 2.0743 lr 1.11e-05 722 tok/s
+step 7950/8500 loss 2.2706 lr 1.10e-05 733 tok/s
+step 7975/8500 loss 2.3401 lr 1.09e-05 731 tok/s
+step 8000/8500 loss 2.3940 lr 1.08e-05 747 tok/s
+  [eval step 8000] val_loss 2.2980 (best 2.2980)
+  [gen] Once upon a time,  in a small house, a little red bird lived in the tree with his family. They all loved to play together under the big tree.
+  [gen] The little girl wanted to  help the little boy. The girl said, "I will help you."
+  [gen] In the dark forest,  there was a little bird with the bird in the tree's nest near a big tree. The big tree looked sad and could not find him to find food for you."
+  saved ckpt/tiny18m2/model_8000.pt
+step 8025/8500 loss 2.1166 lr 1.07e-05 461 tok/s
+step 8050/8500 loss 2.6289 lr 1.07e-05 776 tok/s
+step 8075/8500 loss 2.1662 lr 1.06e-05 735 tok/s
+step 8100/8500 loss 2.5184 lr 1.05e-05 771 tok/s
+step 8125/8500 loss 2.0873 lr 1.05e-05 666 tok/s
+step 8150/8500 loss 2.4317 lr 1.04e-05 649 tok/s
+step 8175/8500 loss 2.1693 lr 1.03e-05 653 tok/s
+step 8200/8500 loss 2.1482 lr 1.03e-05 655 tok/s
+step 8225/8500 loss 2.2751 lr 1.02e-05 648 tok/s
+step 8250/8500 loss 2.5418 lr 1.02e-05 647 tok/s
+step 8275/8500 loss 2.2481 lr 1.02e-05 647 tok/s
+step 8300/8500 loss 2.0614 lr 1.01e-05 648 tok/s
+step 8325/8500 loss 2.4343 lr 1.01e-05 648 tok/s
+step 8350/8500 loss 2.2624 lr 1.01e-05 647 tok/s
+step 8375/8500 loss 2.3551 lr 1.01e-05 648 tok/s
+step 8400/8500 loss 2.3815 lr 1.00e-05 648 tok/s
+step 8425/8500 loss 2.2978 lr 1.00e-05 648 tok/s
+step 8450/8500 loss 2.2452 lr 1.00e-05 647 tok/s
+step 8475/8500 loss 2.1216 lr 1.00e-05 648 tok/s
+step 8500/8500 loss 2.3229 lr 1.00e-05 649 tok/s
+  [eval step 8500] val_loss 2.2966 (best 2.2966)
+  [gen] Once upon a time,  in a small house in the big tree lived in a small park with many animals. The other squirrels liked to play near the tall trees and look at all the pretty flowers in the grass.
+  [gen] The little girl wanted to  help the little dog. The cat sat on a tree and said, "Why are you sad?" The cat looked at the cat and wanted to help.
+  [gen] In the dark forest,  there was a girl who wanted to bring food from the town. The boy saw that the cat had the surprise! The bird said, "I had an ant who always ate it!"
+done