safe_upload: v2_full_metrics.csv

v2_full_metrics/v2_full_metrics.csv

@@ -0,0 +1,7 @@
+model,mean_cell_mention,mean_has_dna_tags,mean_has_reasoning_tags,mean_len_chars,mean_motif_consensus_rate,mean_n_canonical_only,mean_n_cited_tfs,mean_n_grounded,mean_n_hallucinated,mean_n_scan_tfs,mean_prog_coverage,mean_tf_grounded_rate,n_rows,std_len_chars,std_n_canonical_only,std_n_cited_tfs,std_n_grounded,std_n_hallucinated,std_n_scan_tfs
+T1_SFT,0.94,0.0,0.3,1958.86,0.19249084249084247,0.56,4.28,2.68,1.04,306.78,0.275,0.549048618048618,50,563.6534399788579,0.5351635264103861,2.4822570374560327,2.1113029152634635,0.9789790600416334,33.6328946122691
+T1_postRL,0.82,0.0,0.12,1102.44,0.18145693687796882,0.48,10.56,9.8,0.28,300.28,0.17650000000000002,0.7585514703161762,50,360.9119648889463,0.72773621594641,17.130277289057524,17.132425397473646,0.567097875150313,23.666888261873382
+T2_SFT,0.84,0.0,0.38,1526.94,0.11521428571428573,0.46,3.56,2.84,0.26,222.4,0.16083333333333333,0.6180952380952381,50,763.0301543189496,0.7799999999999999,3.020993214159873,2.914515397111499,0.5936328831862332,25.75732905407702
+T2_postRL,0.9,0.0,0.44,1094.86,0.15595041887125222,0.58,10.18,9.4,0.2,222.4,0.20345238095238094,0.7907504995004992,50,443.0542634937621,0.8506468127254694,14.90730022505752,14.942556675482278,0.4,25.75732905407702
+T3_SFT,0.92,0.0,0.5,1587.42,0.34900094764273865,0.16,7.64,7.18,0.3,217.72,0.12259523809523809,0.7794792703150912,50,738.1286091190342,0.36660605559646714,14.51862252419285,14.546051010497658,0.6082762530298219,21.96819519214084
+T3_postRL,0.76,0.0,0.64,972.72,0.3445031596216199,0.16,7.34,6.94,0.24,217.72,0.10964285714285715,0.8012388955582233,50,400.94074574679985,0.463033476111609,12.420322057016074,12.439308662461915,0.5499090833947008,21.96819519214084