Upload evals for thinkcleaned-d12-r11.25-ctx4096
Browse files
experiments/thinkcleaned-d12-r11.25-ctx4096/evals/core.json
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "base_model (step 2362)",
|
| 3 |
+
"step": 2362,
|
| 4 |
+
"bpb": {},
|
| 5 |
+
"core_metric": 0.0721177661677981,
|
| 6 |
+
"core_results": {
|
| 7 |
+
"hellaswag_zeroshot": 0.27663812041282654,
|
| 8 |
+
"jeopardy": 0.0004723665479104966,
|
| 9 |
+
"bigbench_qa_wikidata": 0.06968160718679428,
|
| 10 |
+
"arc_easy": 0.3249158263206482,
|
| 11 |
+
"arc_challenge": 0.22184298932552338,
|
| 12 |
+
"copa": 0.5299999713897705,
|
| 13 |
+
"commonsense_qa": 0.2858312726020813,
|
| 14 |
+
"piqa": 0.5473340153694153,
|
| 15 |
+
"openbook_qa": 0.2540000081062317,
|
| 16 |
+
"lambada_openai": 0.20648165047168732,
|
| 17 |
+
"hellaswag": 0.27912765741348267,
|
| 18 |
+
"winograd": 0.5714285969734192,
|
| 19 |
+
"winogrande": 0.5114443302154541,
|
| 20 |
+
"bigbench_dyck_languages": 0.09100000560283661,
|
| 21 |
+
"agi_eval_lsat_ar": 0.260869562625885,
|
| 22 |
+
"bigbench_cs_algorithms": 0.40984848141670227,
|
| 23 |
+
"bigbench_operators": 0.0476190485060215,
|
| 24 |
+
"bigbench_repeat_copy_logic": 0.0,
|
| 25 |
+
"squad": 0.010879848152399063,
|
| 26 |
+
"coqa": 0.07052486389875412,
|
| 27 |
+
"boolq": 0.5629969239234924,
|
| 28 |
+
"bigbench_language_identification": 0.25849997997283936
|
| 29 |
+
},
|
| 30 |
+
"centered_results": {
|
| 31 |
+
"hellaswag_zeroshot": 0.035517493883768715,
|
| 32 |
+
"jeopardy": 0.0004723665479104966,
|
| 33 |
+
"bigbench_qa_wikidata": 0.06968160718679428,
|
| 34 |
+
"arc_easy": 0.09988776842753093,
|
| 35 |
+
"arc_challenge": -0.03754268089930216,
|
| 36 |
+
"copa": 0.059999942779541016,
|
| 37 |
+
"commonsense_qa": 0.10728909075260161,
|
| 38 |
+
"piqa": 0.09466803073883057,
|
| 39 |
+
"openbook_qa": 0.005333344141642253,
|
| 40 |
+
"lambada_openai": 0.20648165047168732,
|
| 41 |
+
"hellaswag": 0.03883687655131022,
|
| 42 |
+
"winograd": 0.14285719394683838,
|
| 43 |
+
"winogrande": 0.022888660430908203,
|
| 44 |
+
"bigbench_dyck_languages": 0.09100000560283661,
|
| 45 |
+
"agi_eval_lsat_ar": 0.07608695328235625,
|
| 46 |
+
"bigbench_cs_algorithms": 0.40984848141670227,
|
| 47 |
+
"bigbench_operators": 0.0476190485060215,
|
| 48 |
+
"bigbench_repeat_copy_logic": 0.0,
|
| 49 |
+
"squad": 0.010879848152399063,
|
| 50 |
+
"coqa": 0.07052486389875412,
|
| 51 |
+
"boolq": -0.1500080949381778,
|
| 52 |
+
"bigbench_language_identification": 0.18426840481060436
|
| 53 |
+
},
|
| 54 |
+
"conditioned_samples": [],
|
| 55 |
+
"unconditioned_samples": []
|
| 56 |
+
}
|
experiments/thinkcleaned-d12-r11.25-ctx4096/evals/samples.json
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "base_model (step 2362)",
|
| 3 |
+
"step": 2362,
|
| 4 |
+
"bpb": {},
|
| 5 |
+
"core_metric": null,
|
| 6 |
+
"core_results": null,
|
| 7 |
+
"centered_results": null,
|
| 8 |
+
"conditioned_samples": [
|
| 9 |
+
{
|
| 10 |
+
"prompt": "The capital of France is",
|
| 11 |
+
"text": "<|bos|>The capital of France is the capital of the United States, and the capital of the United States is the"
|
| 12 |
+
},
|
| 13 |
+
{
|
| 14 |
+
"prompt": "The chemical symbol of gold is",
|
| 15 |
+
"text": "<|bos|>The chemical symbol of gold is the same as that of silver, and the same as that of gold. The"
|
| 16 |
+
},
|
| 17 |
+
{
|
| 18 |
+
"prompt": "If yesterday was Friday, then tomorrow will be",
|
| 19 |
+
"text": "<|bos|>If yesterday was Friday, then tomorrow will be Friday. The\n\nLord's day is the day of the Lord's\n\nSu"
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"prompt": "The opposite of hot is",
|
| 23 |
+
"text": "<|bos|>The opposite of hot is the same as hot, and the same as hot. The same word is used"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"prompt": "The planets of the solar system are:",
|
| 27 |
+
"text": "<|bos|>The planets of the solar system are: the sun, the moon, the stars, the sun, the moon, the"
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"prompt": "My favorite color is",
|
| 31 |
+
"text": "<|bos|>My favorite color is the same as the \"sweetness of the air,\" and the \"s"
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"prompt": "If 5*x + 3 = 13, then x is",
|
| 35 |
+
"text": "<|bos|>If 5*x + 3 = 13, then x is the number of the feet of the body, and the number of feet of the"
|
| 36 |
+
}
|
| 37 |
+
],
|
| 38 |
+
"unconditioned_samples": [
|
| 39 |
+
"<|bos|>PREFACE.\n\nto the works of an author, who has made us the followers of another and of a different school. It had great influence in the University and the Church; but in the prime of education there was hardly a University teacher-at least no I. KIL.BC was taught in any University.\n\nThe Principles of our Composition.-I desire, therefore, to say the\n\nDedication of the Elements and Elements of a Reading.\n\nI I\n\nTHE FIRST SECTION OF THE GENERAL treatise of the Epistle was written in a peculiar manner, which required only a Bibliography from the author of it to be complete.\n\nThis also bore upon habits of thought and",
|
| 40 |
+
"<|bos|>As iron steamer, double iron Cheston, one of the items of powder in the North Star, to be the first imparted to town, passed his.\n\nHowse rushing with impetuosity as he proceeded! At all points magnetic influence and agency, exercises for a moment; but oh! then afterwards, by mere force of force man becomes not only moral but magnetic.\n\nWolverhampton could have easily enough rendered the country very safe from attack, but on his arrival at Harrisburg, he found the genuine Roman magnet not perfectly made out, and apparently launches along the extremities of the ever-smoothed Western Atlantic.\n\nBok",
|
| 41 |
+
"<|bos|>ady Christians, Sanghahrochs with one or two of her youths and who are the most affectionate, seem to start these fine ladies with a very languid and all eloquent manner, and excite the imaginations of the people, who are somewhat startled by it. No one could be more realistic than Tunghah.\n\nAnd when Miss Jermyn and her daughters heard-or in the light of another spirit felt then the power of such eloquence-they sent a sacred present to the holy being to \"translate\" him. That is not seen in this world, and can find no place in it on earth.\n\nValiantly ladies",
|
| 42 |
+
"<|bos|> fanaticism and fanaticism which has given such deep and varied expression among Christians; and it shows itself in its efforts to produce a future as distinguished from present amelioration as we contemplate this question of crime and punishment, the instrumentality employed in that elevated system of morality which is embodied in their laws and established laws. By considering\n\nit as finally fettered in the dominion of an immaterial, eternal reason, touching it as among the mysteries of the universe, by which mortal men in all civilized countries are regulated and admonished, we are prepared to verify assertors of this vital doctrine a number of false and false facts. I will now only add",
|
| 43 |
+
"<|bos|>PREFACE.\n\nCONTENTS.\n\nvii ...415 599 610 462 151-161 Charles II., King of Spain, and prosecutor of the French, ca. 143-168.\n\n1741794, king of Poland, excommunicated and imprisoned in France and Germany, 18-19\n\n178-190 Charles I., King of France, and His ministers, Gen. John Gen1, 190-191 Charles II., King of France, and His French ministers, Count de M\u00e9di\u00e9re, 192-193 George III.,",
|
| 44 |
+
"<|bos|>PREFACE.\n\nIf any of these places were provided the expedition was understood to have had for its object a simple expedition into the interior, for the capture of which the war is responsible.\n\nThis war without doubt won much greater and purer victories to our country than some of those campaigns experienced in the last two years. The present war is an epoch in the history of this country because in it the great difficulty of driving superior fleets and troops overcomes the ignorance and prejudices of those who have to enter it in time of political transition; it is no war because in overcoming difficulties we have immediately to find and seek (a thing in the simple military",
|
| 45 |
+
"<|bos|>ated have the consideration more than the influence of fortification, which is always the wisest way of awakening and gratifying appetite, and which, if the means be 1\n\ncieved, ameliorates and simplifies the symptoms of a prescribed form of life that it has no warlike sound.\n\nCheops was composed by Sir H. Walpole of Brown's bones; three are believed to be the four grand points in the history of the human race that will never live in aeons, and send certain death to their kind.\n\nThere remains the vascular), while its beauties were so congenial to the southern society to which they belong, that the",
|
| 46 |
+
"<|bos|>PREFACE.\n\nMODERN REPRODUCTION.\n\nThe author has a strong and conscientious predilection for the theory of crude reproduction. He therefore presents, with some of his more popular works, a collection of facts on this branch of moral science. The collected editions of his works have been published privately. Hamilton has laboured during the past sixteen years to establish a system of indirectly controllable morals, which embodies at best a partial abridgment of the first six volumes of moral Biology; and Consent to the institution of Plurality of Worldly Honor, having a strong personal interest in the present agitation for the existing creation of our species, and such an interest as"
|
| 47 |
+
]
|
| 48 |
+
}
|
experiments/thinkcleaned-d12-r11.25-ctx4096/evals/val_bpb.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "base_model (step 2362)",
|
| 3 |
+
"step": 2362,
|
| 4 |
+
"bpb": {
|
| 5 |
+
"val": 1.0433092019869723
|
| 6 |
+
},
|
| 7 |
+
"core_metric": null,
|
| 8 |
+
"core_results": null,
|
| 9 |
+
"centered_results": null,
|
| 10 |
+
"conditioned_samples": [],
|
| 11 |
+
"unconditioned_samples": []
|
| 12 |
+
}
|