===== MB round-1 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool/train/ksb100 weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.009, "harmlessness": 0.013}, "P(pi>mu)": {"helpfulness": 0.496, "harmlessness": 0.507}, "w_nbs": {"helpfulness": 1.0, "harmlessness": 0.0}, "w_ks": {"helpfulness": 0.556, "harmlessness": 0.444}, "worst": "helpfulness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp1] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool/train/ksb100

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  94%|█████████▍| 2121/2250 [00:00<00:00, 20752.89 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 19369.29 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 15327.37 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, 194100.86 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 192568.08 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, 53167.83 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 52326.76 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, 288281.52 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 285490.80 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, 65015.87 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 63739.35 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-2 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp1/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": 0.001, "harmlessness": 0.022}, "P(pi>mu)": {"helpfulness": 0.501, "harmlessness": 0.511}, "w_nbs": {"helpfulness": 0.941, "harmlessness": 0.059}, "w_ks": {"helpfulness": 0.586, "harmlessness": 0.414}, "worst": "helpfulness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp2] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp1/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  94%|█████████▎| 2106/2250 [00:00<00:00, 20016.33 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18674.37 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 11248.04 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, 169919.95 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 168668.73 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, 52376.42 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 51491.65 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, 243402.04 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 241310.83 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, 62567.93 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 61327.41 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-3 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp2/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.009, "harmlessness": 0.036}, "P(pi>mu)": {"helpfulness": 0.496, "harmlessness": 0.518}, "w_nbs": {"helpfulness": 1.0, "harmlessness": 0.0}, "w_ks": {"helpfulness": 0.644, "harmlessness": 0.356}, "worst": "helpfulness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp3] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp2/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  92%|█████████▏| 2073/2250 [00:00<00:00, 19720.73 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18590.38 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 13542.77 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, 165497.85 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 164436.65 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, 54462.99 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 53589.00 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, 232025.77 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 230197.68 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, 62849.20 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 61724.51 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-4 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp3/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": 0.012, "harmlessness": -0.026}, "P(pi>mu)": {"helpfulness": 0.506, "harmlessness": 0.487}, "w_nbs": {"helpfulness": 0.0, "harmlessness": 1.0}, "w_ks": {"helpfulness": 0.451, "harmlessness": 0.549}, "worst": "harmlessness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp4] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp3/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  92%|█████████▏| 2077/2250 [00:00<00:00, 19748.48 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18732.08 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 15029.90 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, 162103.58 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 161123.83 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, 51582.84 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 50793.26 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, 230890.42 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 228996.72 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, 63573.18 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 62441.26 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-5 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp4/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": 0.014, "harmlessness": 0.008}, "P(pi>mu)": {"helpfulness": 0.507, "harmlessness": 0.504}, "w_nbs": {"helpfulness": 0.376, "harmlessness": 0.624}, "w_ks": {"helpfulness": 0.478, "harmlessness": 0.522}, "worst": "harmlessness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp5] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp4/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  91%|█████████ | 2050/2250 [00:00<00:00, 19481.56 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18506.90 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 14515.37 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, 159711.35 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 158720.17 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, 54240.43 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 53349.07 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, 228974.50 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 227062.80 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, 62255.89 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 61134.33 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-6 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp5/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.004, "harmlessness": 0.016}, "P(pi>mu)": {"helpfulness": 0.498, "harmlessness": 0.508}, "w_nbs": {"helpfulness": 1.0, "harmlessness": 0.0}, "w_ks": {"helpfulness": 0.565, "harmlessness": 0.435}, "worst": "helpfulness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp6] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp5/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  90%|█████████ | 2031/2250 [00:00<00:00, 18934.42 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18016.22 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 14591.93 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, 160191.20 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 159113.55 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, 53406.13 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 52510.19 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, 223693.56 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 221873.89 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, 62951.07 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 61753.59 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
===== MB round-7 task=saferlhf anchor=/NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp6/train/dpo weights={"helpfulness":0.5000,"harmlessness":0.5000} =====
manifest 2500
allresp built
scores combined {'pol_42': 2500, 'pol_43': 2500, 'base_44': 2500, 'base_45': 2500}
{"surplus(2P-1)": {"helpfulness": -0.001, "harmlessness": -0.002}, "P(pi>mu)": {"helpfulness": 0.499, "harmlessness": 0.499}, "w_nbs": {"helpfulness": 0.5, "harmlessness": 0.5}, "w_ks": {"helpfulness": 0.5, "harmlessness": 0.5}, "worst": "harmlessness", "n": 5000}
RMPOOL_saferlhf_DONE
[build_bon] 2500 pairs (train 2250/test 250) weights {'helpfulness': 0.5, 'harmlessness': 0.5}
[prep saferlhf EXP=exp/rmpool_mbsfdp7] objs: helpfulness harmlessness parent: /NHNHOME/AIPR/sjkim/gtpref/saferlhf/exp/rmpool_mbsfdp6/train/dpo

Map:   0%|          | 0/2250 [00:00<?, ? examples/s]
Map:  92%|█████████▏| 2077/2250 [00:00<00:00, 19618.09 examples/s]
Map: 100%|██████████| 2250/2250 [00:00<00:00, 18599.69 examples/s]

Map:   0%|          | 0/250 [00:00<?, ? examples/s]
Map: 100%|██████████| 250/250 [00:00<00:00, 14624.29 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, 157820.36 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 156727.41 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, 53232.61 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 52350.27 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, 235370.59 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 2250/2250 [00:00<00:00, 233345.30 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, 60073.10 examples/s]
Saving the dataset (1/1 shards): 100%|██████████| 250/250 [00:00<00:00, 59054.74 examples/s]
[prep] pre_pref + pre_htobj built
dpo htmnpo_harmlessness htmnpo_helpfulness inpo ipo ks maxmin nbs simpo unif 
TRAIN_DONE aiprlab-polymer
{
  "label": "qwen32_saferlhf_dpo",
  "n_objectives": 2,
  "surplus": {
    "helpfulness": 0.3103,
    "harmlessness": 0.3741
  },
  "hoeffding_margin": {
    "helpfulness": 0.0331,
    "harmlessness": 0.0331
  },
  "worst_surplus": 0.3103,
  "individual_rationality": true,
  "nash_welfare_logsum": -2.1535
}
EVAL_saferlhf_DONE aiprlab-polymer
MB_STAGES_DONE