===== MB round-1 task=hs2gemma anchor=/NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/ weights={"helpfulness":0.2500,"correctness":0.2500,"coherence":0.2500,"conciseness":0.2500} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.035, "correctness": -0.033, "coherence": -0.028, "conciseness": 0.001}, "P(pi>mu)": {"helpfulness": 0.482, "correctness": 0.484, "coherence": 0.486, "conciseness": 0.5}, "w_nbs": {"helpfulness": 0.333, "correctness": 0.333, "coherence": 0.333, "conciseness": 0.001}, "w_ks": {"helpfulness": 0.251, "correctness": 0.251, "coherence": 0.251, "conciseness": 0.247}, "worst": "helpfulness", "n": 5000}
RMPOOL_hs2gemma_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.25, 'correctness': 0.25, 'coherence': 0.25, 'conciseness': 0.25}
[prep hs2gemma EXP=exp/rmpool_mbgsp1] objs: helpfulness correctness coherence conciseness parent: /NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  89%|████████▉ | 2000/2250 [00:00<00:00, 16907.03 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 16450.97 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 13273.11 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 187067.55 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 185672.65 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 50874.58 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 50075.26 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 259519.96 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 257165.00 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 56254.08 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 55304.64 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_coherence htmnpo_conciseness htmnpo_correctness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
MB1_TRAIN_FAIL
===== MB round-1 task=hs2gemma anchor=/NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/ weights={"helpfulness":0.2500,"correctness":0.2500,"coherence":0.2500,"conciseness":0.2500} =====
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.035, "correctness": -0.033, "coherence": -0.028, "conciseness": 0.001}, "P(pi>mu)": {"helpfulness": 0.482, "correctness": 0.484, "coherence": 0.486, "conciseness": 0.5}, "w_nbs": {"helpfulness": 0.333, "correctness": 0.333, "coherence": 0.333, "conciseness": 0.001}, "w_ks": {"helpfulness": 0.251, "correctness": 0.251, "coherence": 0.251, "conciseness": 0.247}, "worst": "helpfulness", "n": 5000}
RMPOOL_hs2gemma_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.25, 'correctness': 0.25, 'coherence': 0.25, 'conciseness': 0.25}
[prep hs2gemma EXP=exp/rmpool_mbgsp1] objs: helpfulness correctness coherence conciseness parent: /NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 89581.90 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 89109.06 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24353.21 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24149.61 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 111455.78 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 110953.90 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24815.43 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24626.02 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_coherence htmnpo_conciseness htmnpo_correctness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
MB1_TRAIN_FAIL
===== MB round-1 task=hs2gemma anchor=/NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/ weights={"helpfulness":0.2500,"correctness":0.2500,"coherence":0.2500,"conciseness":0.2500} =====
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.035, "correctness": -0.033, "coherence": -0.028, "conciseness": 0.001}, "P(pi>mu)": {"helpfulness": 0.482, "correctness": 0.484, "coherence": 0.486, "conciseness": 0.5}, "w_nbs": {"helpfulness": 0.333, "correctness": 0.333, "coherence": 0.333, "conciseness": 0.001}, "w_ks": {"helpfulness": 0.251, "correctness": 0.251, "coherence": 0.251, "conciseness": 0.247}, "worst": "helpfulness", "n": 5000}
RMPOOL_hs2gemma_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.25, 'correctness': 0.25, 'coherence': 0.25, 'conciseness': 0.25}
[prep hs2gemma EXP=exp/rmpool_mbgsp1] objs: helpfulness correctness coherence conciseness parent: /NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 96494.72 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 95961.93 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 23385.36 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 23208.34 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 97140.34 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 96503.60 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24685.15 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24469.14 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_coherence htmnpo_conciseness htmnpo_correctness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
MB1_TRAIN_FAIL
===== MB round-1 task=hs2gemma anchor=/NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/ weights={"helpfulness":0.2500,"correctness":0.2500,"coherence":0.2500,"conciseness":0.2500} =====
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.035, "correctness": -0.033, "coherence": -0.028, "conciseness": 0.001}, "P(pi>mu)": {"helpfulness": 0.482, "correctness": 0.484, "coherence": 0.486, "conciseness": 0.5}, "w_nbs": {"helpfulness": 0.333, "correctness": 0.333, "coherence": 0.333, "conciseness": 0.001}, "w_ks": {"helpfulness": 0.251, "correctness": 0.251, "coherence": 0.251, "conciseness": 0.247}, "worst": "helpfulness", "n": 5000}
RMPOOL_hs2gemma_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.25, 'correctness': 0.25, 'coherence': 0.25, 'conciseness': 0.25}
[prep hs2gemma EXP=exp/rmpool_mbgsp1] objs: helpfulness correctness coherence conciseness parent: /NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 94011.77 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 93529.14 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24921.59 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24725.32 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 98320.39 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 97959.10 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24772.05 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24585.03 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_coherence htmnpo_conciseness htmnpo_correctness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
MB1_TRAIN_FAIL
===== MB round-1 task=hs2gemma anchor=/NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/ weights={"helpfulness":0.2500,"correctness":0.2500,"coherence":0.2500,"conciseness":0.2500} =====
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.035, "correctness": -0.033, "coherence": -0.028, "conciseness": 0.001}, "P(pi>mu)": {"helpfulness": 0.482, "correctness": 0.484, "coherence": 0.486, "conciseness": 0.5}, "w_nbs": {"helpfulness": 0.333, "correctness": 0.333, "coherence": 0.333, "conciseness": 0.001}, "w_ks": {"helpfulness": 0.251, "correctness": 0.251, "coherence": 0.251, "conciseness": 0.247}, "worst": "helpfulness", "n": 5000}
RMPOOL_hs2gemma_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.25, 'correctness': 0.25, 'coherence': 0.25, 'conciseness': 0.25}
[prep hs2gemma EXP=exp/rmpool_mbgsp1] objs: helpfulness correctness coherence conciseness parent: /NHNHOME/AIPR/sjkim/policy_bases/models--google--gemma-2-9b-it/snapshots/11c9b309abf73637e4b6f9a3fa1e92e615547819/

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 90595.81 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 90172.51 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24278.21 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 24052.11 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/2250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 101456.55 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 101061.07 examples/s]

Saving the dataset (0/1 shards):   0%|          | 0/250 [00:00<?, ? examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 25731.93 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 25523.98 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_coherence htmnpo_conciseness htmnpo_correctness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif