Text Generation
Transformers
bloom
Eval Results
TheBloke commited on
Commit
d090b24
1 Parent(s): 9290efc

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +1127 -0
README.md ADDED
@@ -0,0 +1,1127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - bigscience/xP3
4
+ license: bigscience-bloom-rail-1.0
5
+ language:
6
+ - ak
7
+ - ar
8
+ - as
9
+ - bm
10
+ - bn
11
+ - ca
12
+ - code
13
+ - en
14
+ - es
15
+ - eu
16
+ - fon
17
+ - fr
18
+ - gu
19
+ - hi
20
+ - id
21
+ - ig
22
+ - ki
23
+ - kn
24
+ - lg
25
+ - ln
26
+ - ml
27
+ - mr
28
+ - ne
29
+ - nso
30
+ - ny
31
+ - or
32
+ - pa
33
+ - pt
34
+ - rn
35
+ - rw
36
+ - sn
37
+ - st
38
+ - sw
39
+ - ta
40
+ - te
41
+ - tn
42
+ - ts
43
+ - tum
44
+ - tw
45
+ - ur
46
+ - vi
47
+ - wo
48
+ - xh
49
+ - yo
50
+ - zh
51
+ - zu
52
+ programming_language:
53
+ - C
54
+ - C++
55
+ - C#
56
+ - Go
57
+ - Java
58
+ - JavaScript
59
+ - Lua
60
+ - PHP
61
+ - Python
62
+ - Ruby
63
+ - Rust
64
+ - Scala
65
+ - TypeScript
66
+ pipeline_tag: text-generation
67
+ inference: false
68
+ widget:
69
+ - text: "一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。Would you rate the previous review as positive, neutral or negative?"
70
+ example_title: "zh-en sentiment"
71
+ - text: "一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?"
72
+ example_title: "zh-zh sentiment"
73
+ - text: "Suggest at least five related search terms to \"Mạng neural nhân tạo\"."
74
+ example_title: "vi-en query"
75
+ - text: "Proposez au moins cinq mots clés concernant «Réseau de neurones artificiels»."
76
+ example_title: "fr-fr query"
77
+ - text: "Explain in a sentence in Telugu what is backpropagation in neural networks."
78
+ example_title: "te-en qa"
79
+ - text: "Why is the sky blue?"
80
+ example_title: "en-en qa"
81
+ - text: "Explain to me in Traditional Chinese what is the difference between Bitcoin and Ethereum."
82
+ example_title: "zh-en qa"
83
+ - text: "Write a code snippet with O(log(n)) computational complexity."
84
+ example_title: "code-en"
85
+ - text: "Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is \"Heroes Come in All Shapes and Sizes\". Story (in Spanish):"
86
+ example_title: "es-en fable"
87
+ - text: "Write a fable about wood elves living in a forest that is suddenly invaded by ogres. The fable is a masterpiece that has achieved praise worldwide and its moral is \"Violence is the last refuge of the incompetent\". Fable (in Hindi):"
88
+ example_title: "hi-en fable"
89
+ - text: "How many sides does a rectangle and heptagon have, when
90
+ combined? Answer this question with some math.
91
+ Ein Rechteck hat 4 Seiten. Ein Siebeneck hat 7 Seiten.
92
+ In Kombination haben sie 4 + 7 = 11 Seiten.
93
+ كم عدد الأضلاع التي يجمعها المربع والمثلث؟
94
+ Répondez à cette question en chinois."
95
+ example_title: "en-de-ar-fr-zh math"
96
+ model-index:
97
+ - name: bloomz
98
+ results:
99
+ - task:
100
+ type: Coreference resolution
101
+ dataset:
102
+ type: winogrande
103
+ name: Winogrande XL (xl)
104
+ config: xl
105
+ split: validation
106
+ revision: a80f460359d1e9a67c006011c94de42a8759430c
107
+ metrics:
108
+ - type: Accuracy
109
+ value: 59.27
110
+ - task:
111
+ type: Coreference resolution
112
+ dataset:
113
+ type: Muennighoff/xwinograd
114
+ name: XWinograd (en)
115
+ config: en
116
+ split: test
117
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
118
+ metrics:
119
+ - type: Accuracy
120
+ value: 69.08
121
+ - task:
122
+ type: Coreference resolution
123
+ dataset:
124
+ type: Muennighoff/xwinograd
125
+ name: XWinograd (fr)
126
+ config: fr
127
+ split: test
128
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
129
+ metrics:
130
+ - type: Accuracy
131
+ value: 68.67
132
+ - task:
133
+ type: Coreference resolution
134
+ dataset:
135
+ type: Muennighoff/xwinograd
136
+ name: XWinograd (jp)
137
+ config: jp
138
+ split: test
139
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
140
+ metrics:
141
+ - type: Accuracy
142
+ value: 59.65
143
+ - task:
144
+ type: Coreference resolution
145
+ dataset:
146
+ type: Muennighoff/xwinograd
147
+ name: XWinograd (pt)
148
+ config: pt
149
+ split: test
150
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
151
+ metrics:
152
+ - type: Accuracy
153
+ value: 64.26
154
+ - task:
155
+ type: Coreference resolution
156
+ dataset:
157
+ type: Muennighoff/xwinograd
158
+ name: XWinograd (ru)
159
+ config: ru
160
+ split: test
161
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
162
+ metrics:
163
+ - type: Accuracy
164
+ value: 60.95
165
+ - task:
166
+ type: Coreference resolution
167
+ dataset:
168
+ type: Muennighoff/xwinograd
169
+ name: XWinograd (zh)
170
+ config: zh
171
+ split: test
172
+ revision: 9dd5ea5505fad86b7bedad667955577815300cee
173
+ metrics:
174
+ - type: Accuracy
175
+ value: 70.24
176
+ - task:
177
+ type: Natural language inference
178
+ dataset:
179
+ type: anli
180
+ name: ANLI (r1)
181
+ config: r1
182
+ split: validation
183
+ revision: 9dbd830a06fea8b1c49d6e5ef2004a08d9f45094
184
+ metrics:
185
+ - type: Accuracy
186
+ value: 48.6
187
+ - task:
188
+ type: Natural language inference
189
+ dataset:
190
+ type: anli
191
+ name: ANLI (r2)
192
+ config: r2
193
+ split: validation
194
+ revision: 9dbd830a06fea8b1c49d6e5ef2004a08d9f45094
195
+ metrics:
196
+ - type: Accuracy
197
+ value: 44.1
198
+ - task:
199
+ type: Natural language inference
200
+ dataset:
201
+ type: anli
202
+ name: ANLI (r3)
203
+ config: r3
204
+ split: validation
205
+ revision: 9dbd830a06fea8b1c49d6e5ef2004a08d9f45094
206
+ metrics:
207
+ - type: Accuracy
208
+ value: 45.5
209
+ - task:
210
+ type: Natural language inference
211
+ dataset:
212
+ type: super_glue
213
+ name: SuperGLUE (cb)
214
+ config: cb
215
+ split: validation
216
+ revision: 9e12063561e7e6c79099feb6d5a493142584e9e2
217
+ metrics:
218
+ - type: Accuracy
219
+ value: 82.14
220
+ - task:
221
+ type: Natural language inference
222
+ dataset:
223
+ type: super_glue
224
+ name: SuperGLUE (rte)
225
+ config: rte
226
+ split: validation
227
+ revision: 9e12063561e7e6c79099feb6d5a493142584e9e2
228
+ metrics:
229
+ - type: Accuracy
230
+ value: 85.56
231
+ - task:
232
+ type: Natural language inference
233
+ dataset:
234
+ type: xnli
235
+ name: XNLI (ar)
236
+ config: ar
237
+ split: validation
238
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
239
+ metrics:
240
+ - type: Accuracy
241
+ value: 60.68
242
+ - task:
243
+ type: Natural language inference
244
+ dataset:
245
+ type: xnli
246
+ name: XNLI (bg)
247
+ config: bg
248
+ split: validation
249
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
250
+ metrics:
251
+ - type: Accuracy
252
+ value: 48.43
253
+ - task:
254
+ type: Natural language inference
255
+ dataset:
256
+ type: xnli
257
+ name: XNLI (de)
258
+ config: de
259
+ split: validation
260
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
261
+ metrics:
262
+ - type: Accuracy
263
+ value: 54.38
264
+ - task:
265
+ type: Natural language inference
266
+ dataset:
267
+ type: xnli
268
+ name: XNLI (el)
269
+ config: el
270
+ split: validation
271
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
272
+ metrics:
273
+ - type: Accuracy
274
+ value: 47.43
275
+ - task:
276
+ type: Natural language inference
277
+ dataset:
278
+ type: xnli
279
+ name: XNLI (en)
280
+ config: en
281
+ split: validation
282
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
283
+ metrics:
284
+ - type: Accuracy
285
+ value: 67.47
286
+ - task:
287
+ type: Natural language inference
288
+ dataset:
289
+ type: xnli
290
+ name: XNLI (es)
291
+ config: es
292
+ split: validation
293
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
294
+ metrics:
295
+ - type: Accuracy
296
+ value: 61.24
297
+ - task:
298
+ type: Natural language inference
299
+ dataset:
300
+ type: xnli
301
+ name: XNLI (fr)
302
+ config: fr
303
+ split: validation
304
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
305
+ metrics:
306
+ - type: Accuracy
307
+ value: 61.37
308
+ - task:
309
+ type: Natural language inference
310
+ dataset:
311
+ type: xnli
312
+ name: XNLI (hi)
313
+ config: hi
314
+ split: validation
315
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
316
+ metrics:
317
+ - type: Accuracy
318
+ value: 60.2
319
+ - task:
320
+ type: Natural language inference
321
+ dataset:
322
+ type: xnli
323
+ name: XNLI (ru)
324
+ config: ru
325
+ split: validation
326
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
327
+ metrics:
328
+ - type: Accuracy
329
+ value: 54.02
330
+ - task:
331
+ type: Natural language inference
332
+ dataset:
333
+ type: xnli
334
+ name: XNLI (sw)
335
+ config: sw
336
+ split: validation
337
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
338
+ metrics:
339
+ - type: Accuracy
340
+ value: 52.09
341
+ - task:
342
+ type: Natural language inference
343
+ dataset:
344
+ type: xnli
345
+ name: XNLI (th)
346
+ config: th
347
+ split: validation
348
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
349
+ metrics:
350
+ - type: Accuracy
351
+ value: 43.78
352
+ - task:
353
+ type: Natural language inference
354
+ dataset:
355
+ type: xnli
356
+ name: XNLI (tr)
357
+ config: tr
358
+ split: validation
359
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
360
+ metrics:
361
+ - type: Accuracy
362
+ value: 45.7
363
+ - task:
364
+ type: Natural language inference
365
+ dataset:
366
+ type: xnli
367
+ name: XNLI (ur)
368
+ config: ur
369
+ split: validation
370
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
371
+ metrics:
372
+ - type: Accuracy
373
+ value: 50.8
374
+ - task:
375
+ type: Natural language inference
376
+ dataset:
377
+ type: xnli
378
+ name: XNLI (vi)
379
+ config: vi
380
+ split: validation
381
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
382
+ metrics:
383
+ - type: Accuracy
384
+ value: 61.0
385
+ - task:
386
+ type: Natural language inference
387
+ dataset:
388
+ type: xnli
389
+ name: XNLI (zh)
390
+ config: zh
391
+ split: validation
392
+ revision: a5a45e4ff92d5d3f34de70aaf4b72c3bdf9f7f16
393
+ metrics:
394
+ - type: Accuracy
395
+ value: 56.91
396
+ - task:
397
+ type: Program synthesis
398
+ dataset:
399
+ type: openai_humaneval
400
+ name: HumanEval
401
+ config: None
402
+ split: test
403
+ revision: e8dc562f5de170c54b5481011dd9f4fa04845771
404
+ metrics:
405
+ - type: Pass@1
406
+ value: 12.06
407
+ - type: Pass@10
408
+ value: 26.53
409
+ - type: Pass@100
410
+ value: 48.44
411
+ - task:
412
+ type: Sentence completion
413
+ dataset:
414
+ type: story_cloze
415
+ name: StoryCloze (2016)
416
+ config: "2016"
417
+ split: validation
418
+ revision: e724c6f8cdf7c7a2fb229d862226e15b023ee4db
419
+ metrics:
420
+ - type: Accuracy
421
+ value: 96.26
422
+ - task:
423
+ type: Sentence completion
424
+ dataset:
425
+ type: super_glue
426
+ name: SuperGLUE (copa)
427
+ config: copa
428
+ split: validation
429
+ revision: 9e12063561e7e6c79099feb6d5a493142584e9e2
430
+ metrics:
431
+ - type: Accuracy
432
+ value: 91.0
433
+ - task:
434
+ type: Sentence completion
435
+ dataset:
436
+ type: xcopa
437
+ name: XCOPA (et)
438
+ config: et
439
+ split: validation
440
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
441
+ metrics:
442
+ - type: Accuracy
443
+ value: 51.0
444
+ - task:
445
+ type: Sentence completion
446
+ dataset:
447
+ type: xcopa
448
+ name: XCOPA (ht)
449
+ config: ht
450
+ split: validation
451
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
452
+ metrics:
453
+ - type: Accuracy
454
+ value: 58.0
455
+ - task:
456
+ type: Sentence completion
457
+ dataset:
458
+ type: xcopa
459
+ name: XCOPA (id)
460
+ config: id
461
+ split: validation
462
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
463
+ metrics:
464
+ - type: Accuracy
465
+ value: 86.0
466
+ - task:
467
+ type: Sentence completion
468
+ dataset:
469
+ type: xcopa
470
+ name: XCOPA (it)
471
+ config: it
472
+ split: validation
473
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
474
+ metrics:
475
+ - type: Accuracy
476
+ value: 74.0
477
+ - task:
478
+ type: Sentence completion
479
+ dataset:
480
+ type: xcopa
481
+ name: XCOPA (qu)
482
+ config: qu
483
+ split: validation
484
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
485
+ metrics:
486
+ - type: Accuracy
487
+ value: 56.0
488
+ - task:
489
+ type: Sentence completion
490
+ dataset:
491
+ type: xcopa
492
+ name: XCOPA (sw)
493
+ config: sw
494
+ split: validation
495
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
496
+ metrics:
497
+ - type: Accuracy
498
+ value: 64.0
499
+ - task:
500
+ type: Sentence completion
501
+ dataset:
502
+ type: xcopa
503
+ name: XCOPA (ta)
504
+ config: ta
505
+ split: validation
506
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
507
+ metrics:
508
+ - type: Accuracy
509
+ value: 69.0
510
+ - task:
511
+ type: Sentence completion
512
+ dataset:
513
+ type: xcopa
514
+ name: XCOPA (th)
515
+ config: th
516
+ split: validation
517
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
518
+ metrics:
519
+ - type: Accuracy
520
+ value: 58.0
521
+ - task:
522
+ type: Sentence completion
523
+ dataset:
524
+ type: xcopa
525
+ name: XCOPA (tr)
526
+ config: tr
527
+ split: validation
528
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
529
+ metrics:
530
+ - type: Accuracy
531
+ value: 57.0
532
+ - task:
533
+ type: Sentence completion
534
+ dataset:
535
+ type: xcopa
536
+ name: XCOPA (vi)
537
+ config: vi
538
+ split: validation
539
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
540
+ metrics:
541
+ - type: Accuracy
542
+ value: 87.0
543
+ - task:
544
+ type: Sentence completion
545
+ dataset:
546
+ type: xcopa
547
+ name: XCOPA (zh)
548
+ config: zh
549
+ split: validation
550
+ revision: 37f73c60fb123111fa5af5f9b705d0b3747fd187
551
+ metrics:
552
+ - type: Accuracy
553
+ value: 90.0
554
+ - task:
555
+ type: Sentence completion
556
+ dataset:
557
+ type: Muennighoff/xstory_cloze
558
+ name: XStoryCloze (ar)
559
+ config: ar
560
+ split: validation
561
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
562
+ metrics:
563
+ - type: Accuracy
564
+ value: 92.79
565
+ - task:
566
+ type: Sentence completion
567
+ dataset:
568
+ type: Muennighoff/xstory_cloze
569
+ name: XStoryCloze (es)
570
+ config: es
571
+ split: validation
572
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
573
+ metrics:
574
+ - type: Accuracy
575
+ value: 94.37
576
+ - task:
577
+ type: Sentence completion
578
+ dataset:
579
+ type: Muennighoff/xstory_cloze
580
+ name: XStoryCloze (eu)
581
+ config: eu
582
+ split: validation
583
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
584
+ metrics:
585
+ - type: Accuracy
586
+ value: 86.9
587
+ - task:
588
+ type: Sentence completion
589
+ dataset:
590
+ type: Muennighoff/xstory_cloze
591
+ name: XStoryCloze (hi)
592
+ config: hi
593
+ split: validation
594
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
595
+ metrics:
596
+ - type: Accuracy
597
+ value: 88.42
598
+ - task:
599
+ type: Sentence completion
600
+ dataset:
601
+ type: Muennighoff/xstory_cloze
602
+ name: XStoryCloze (id)
603
+ config: id
604
+ split: validation
605
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
606
+ metrics:
607
+ - type: Accuracy
608
+ value: 92.12
609
+ - task:
610
+ type: Sentence completion
611
+ dataset:
612
+ type: Muennighoff/xstory_cloze
613
+ name: XStoryCloze (my)
614
+ config: my
615
+ split: validation
616
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
617
+ metrics:
618
+ - type: Accuracy
619
+ value: 52.35
620
+ - task:
621
+ type: Sentence completion
622
+ dataset:
623
+ type: Muennighoff/xstory_cloze
624
+ name: XStoryCloze (ru)
625
+ config: ru
626
+ split: validation
627
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
628
+ metrics:
629
+ - type: Accuracy
630
+ value: 81.73
631
+ - task:
632
+ type: Sentence completion
633
+ dataset:
634
+ type: Muennighoff/xstory_cloze
635
+ name: XStoryCloze (sw)
636
+ config: sw
637
+ split: validation
638
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
639
+ metrics:
640
+ - type: Accuracy
641
+ value: 79.81
642
+ - task:
643
+ type: Sentence completion
644
+ dataset:
645
+ type: Muennighoff/xstory_cloze
646
+ name: XStoryCloze (te)
647
+ config: te
648
+ split: validation
649
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
650
+ metrics:
651
+ - type: Accuracy
652
+ value: 81.2
653
+ - task:
654
+ type: Sentence completion
655
+ dataset:
656
+ type: Muennighoff/xstory_cloze
657
+ name: XStoryCloze (zh)
658
+ config: zh
659
+ split: validation
660
+ revision: 8bb76e594b68147f1a430e86829d07189622b90d
661
+ metrics:
662
+ - type: Accuracy
663
+ value: 93.12
664
+ ---
665
+
666
+ <!-- header start -->
667
+ <div style="width: 100%;">
668
+ <img src="https://i.imgur.com/EBdldam.jpg" alt="TheBlokeAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
669
+ </div>
670
+ <div style="display: flex; justify-content: space-between; width: 100%;">
671
+ <div style="display: flex; flex-direction: column; align-items: flex-start;">
672
+ <p><a href="https://discord.gg/theblokeai">Chat & support: my new Discord server</a></p>
673
+ </div>
674
+ <div style="display: flex; flex-direction: column; align-items: flex-end;">
675
+ <p><a href="https://www.patreon.com/TheBlokeAI">Want to contribute? TheBloke's Patreon page</a></p>
676
+ </div>
677
+ </div>
678
+ <!-- header end -->
679
+
680
+ # BigScience's BLOOMZ 176B GPTQ
681
+
682
+ These files are GPTQ 4bit model files for [BigScience's BLOOMZ](https://huggingface.co/bigscience/bloomz).
683
+
684
+ It is the result of quantising to 4bit using [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ).
685
+
686
+ **This is a BIG model! 2 x 80GB or 3 x 48GB GPUs are required**
687
+
688
+ ## Important note: files must be joined before use
689
+
690
+ It is not currently possible to shard GPTQ files, therefore the model file is one single 100+GB file.
691
+
692
+ Huggingface Hub has a 50GB per-file limit. I have therefore been forced to split the file in to three parts for upload.
693
+
694
+ I did this using the simple *nix command `split`.
695
+
696
+ To join the files on any *nix system, run:
697
+ ```
698
+ cat gptq_model-4bit--1g.split* > gptq_model-4bit--1g.safetensors
699
+ ```
700
+
701
+ To join the files on Windows, open a Command Prompt and run:
702
+ ```
703
+ COPY /B gptq_model-4bit--1g.splitaa + gptq_model-4bit--1g.splitab + gptq_model-4bit--1g.splitac gptq_model-4bit--1g.safetensors
704
+ ```
705
+
706
+ The SHA256SUM of the joined file will be:
707
+
708
+ Once you have the joined file, you can safely delete `gptq_model-4bit--1g.split*`.
709
+
710
+ ## Repositories available
711
+
712
+ * [4-bit GPTQ model for GPU inference](https://huggingface.co/TheBlokeAI/bloomz-175B-GPTQ)
713
+ * [Unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/bigscience/bloomz)
714
+
715
+ ## Two files provided - separate branches
716
+
717
+ - Main branch:
718
+ - Group Size = None
719
+ - Desc Act (act-order) = True
720
+ - This version will use the least possible VRAM, and should have higher inference performance in CUDA mode
721
+ - Branch `group_size_128g`:
722
+ - Group Size = 128g
723
+ - Desc Act (act-oder) = True
724
+ - This version will use more VRAM, which shouldn't be a problem as it shouldn't exceed 2 x 80GB or 3 x 48GB cards.
725
+ - However CUDA inference performance is likely to be a lot slower, possibly necessitating the use of Triton mode.
726
+
727
+ By default you will download the first file, unless you choose to download from branch `group_size_128g`.
728
+
729
+ ## Prompt template: none
730
+
731
+ ```
732
+ Translate to English: Je t’aime.
733
+ Translation:
734
+ ```
735
+
736
+ ## How to easily download and use this model in text-generation-webui
737
+
738
+ Please make sure you're using the latest version of text-generation-webui.
739
+
740
+ Note 1: this is a non-Llama model which cannot be used with ExLlama. Use Loader: AutoGPTQ.
741
+
742
+ Note 2: As described above, you must join the files after downloading and before loading in text-generation-webui.
743
+
744
+ 1. Click the **Model tab**.
745
+ 2. Under **Download custom model or LoRA**, enter `TheBloke/bloomz-176B-GPTQ`.
746
+ - If you would rather download the group_size 128g version, enter `TheBloke/bloomz-176B-GPTQ:group_size_128g`
747
+ 3. Click **Download**.
748
+ 4. The model will start downloading. Once it's finished it will say "Done". This is a huge model so it may take a while!
749
+ 5. Now follow the steps described above to join the model to get a single `.safetensors` file.
750
+ 6. Untick **Autoload model**
751
+ 7. In the top left, click the refresh icon next to **Model**.
752
+ 8. In the **Model** dropdown, choose the model you just downloaded: `bloomz-176B-GPTQ`
753
+ 9. Make sure Loader is set to AutGPTQ.
754
+ 10. This model cannot load on one GPU, so you should set **GPU Memory** accordingly.
755
+ - If using two 80GB GPUs, try: GPU0 = 60GB, GPU1 = 79GB
756
+ - If using three 48GB GPUs, try: GPU0 = 30GB, GPU1 = 47GB, GPU2 = 47GB
757
+ 11. Click **Save settings** to save your settings, and then **Reload** to load the model.
758
+ 12. The model will load, and is now ready for use!
759
+ 13. Once you're ready, click the **Text Generation tab** and enter a prompt to get started!
760
+
761
+ ## How to use this GPTQ model from Python code
762
+
763
+ First make sure you have [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ) installed:
764
+
765
+ `GITHUB_ACTIONS=true pip install auto-gptq`
766
+
767
+ Because this model has to be joined locally, you must first download it. Example download code:
768
+
769
+ ```python
770
+ from huggingface_hub import snapshot_download
771
+ snapshot_download(repo_id="TheBloke/bloomz-176B-GPTQ",
772
+ local_dir="/workspace/models/bloomz-176GB-GPTQ",
773
+ local_dir_use_symlinks=False)
774
+ ```
775
+
776
+ If you want to download the group_size 128g file instead, add `revision="group_size_128g"` to the above command.
777
+
778
+ Now join the three `split` files up as described above, to get a single `safetensors` file.
779
+
780
+ Then try the following example code:
781
+
782
+ ```python
783
+ from transformers import AutoTokenizer, pipeline, logging
784
+ from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
785
+ import argparse
786
+
787
+ # Use the local path you downloaded the model to and joined the split files in
788
+ model_name_or_path = "/workspace/models/bloomz-176GB-GPTQ"
789
+ model_basename = "gptq_model-4bit--1g"
790
+
791
+ use_triton = False
792
+
793
+ tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)
794
+
795
+ model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,
796
+ model_basename=model_basename,
797
+ use_safetensors=True,
798
+ trust_remote_code=False,
799
+ device="cuda:0",
800
+ use_triton=use_triton,
801
+ quantize_config=None)
802
+
803
+ prompt = "Translate this to French: AI is the future of computing"
804
+ prompt_template=f'''{prompt}
805
+ Translation:
806
+ '''
807
+
808
+ print("\n\n*** Generate:")
809
+
810
+ input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
811
+ output = model.generate(inputs=input_ids, temperature=0.7, max_new_tokens=512)
812
+ print(tokenizer.decode(output[0]))
813
+
814
+ # Inference can also be done using transformers' pipeline
815
+
816
+ # Prevent printing spurious transformers error when using pipeline with AutoGPTQ
817
+ logging.set_verbosity(logging.CRITICAL)
818
+
819
+ print("*** Pipeline:")
820
+ pipe = pipeline(
821
+ "text-generation",
822
+ model=model,
823
+ tokenizer=tokenizer,
824
+ max_new_tokens=512,
825
+ temperature=0.7,
826
+ top_p=0.95,
827
+ repetition_penalty=1.15
828
+ )
829
+
830
+ print(pipe(prompt_template)[0]['generated_text'])
831
+ ```
832
+
833
+ ## Provided files
834
+
835
+ ## Main branch:
836
+
837
+ **gptq_model-4bit--1g.safetensors**
838
+
839
+ This will work with AutoGPTQ. It is untested with GPTQ-for-LLaMa. It will *not* work with ExLlama.
840
+
841
+ It was created with group_size none (-1) to reduce VRAM usage, and with --act-order (desc_act) to increase inference speed.
842
+
843
+ * `gptq_model-4bit-128g.safetensors`
844
+ * Works with AutoGPTQ in CUDA or Triton modes.
845
+ * Does NOT work with [ExLlama](https://github.com/turboderp/exllama) as it's not a Llama model.
846
+ * Works with GPTQ-for-LLaMa in CUDA mode. May have issues with GPTQ-for-LLaMa Triton mode.
847
+ * Works with text-generation-webui, including one-click-installers.
848
+ * Parameters: Groupsize = -1. Act Order / desc_act = True.
849
+
850
+ ## Branch `group_size_128g`
851
+
852
+ **gptq_model-4bit-128g.safetensors**
853
+
854
+ This will work with AutoGPTQ. It is untested with GPTQ-for-LLaMa. It will *not* work with ExLlama.
855
+
856
+ It was created with both group_size 128g and --act-order (desc_act) for increased inference quality.
857
+
858
+ **Note** Using group_size + desc_act together can significantly lower performance in AutoGPTQ CUDA. You might want to try AutoGPTQ Triton mode instead (Linux only.)
859
+
860
+ * `gptq_model-4bit-128g.safetensors`
861
+ * Works with AutoGPTQ in CUDA or Triton modes.
862
+ * Does NOT work with [ExLlama](https://github.com/turboderp/exllama) as it's not a Llama model.
863
+ * Untested with GPTQ-for-LLaMa.
864
+ * Works with text-generation-webui, including one-click-installers.
865
+ * Parameters: Groupsize = 128. Act Order / desc_act = True.
866
+
867
+ <!-- footer start -->
868
+ ## Discord
869
+
870
+ For further support, and discussions on these models and AI in general, join us at:
871
+
872
+ [TheBloke AI's Discord server](https://discord.gg/theblokeai)
873
+
874
+ ## Thanks, and how to contribute.
875
+
876
+ Thanks to the [chirper.ai](https://chirper.ai) team!
877
+
878
+ I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.
879
+
880
+ If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.
881
+
882
+ Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.
883
+
884
+ * Patreon: https://patreon.com/TheBlokeAI
885
+ * Ko-Fi: https://ko-fi.com/TheBlokeAI
886
+
887
+ **Special thanks to**: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.
888
+
889
+ **Patreon special mentions**: zynix , ya boyyy, Trenton Dambrowitz, Imad Khwaja, Alps Aficionado, chris gileta, John Detwiler, Willem Michiel, RoA, Mano Prime, Rainer Wilmers, Fred von Graf, Matthew Berman, Ghost , Nathan LeClaire, Iucharbius , Ai Maven, Illia Dulskyi, Joseph William Delisle, Space Cruiser, Lone Striker, Karl Bernard, Eugene Pentland, Greatston Gnanesh, Jonathan Leane, Randy H, Pierre Kircher, Willian Hasse, Stephen Murray, Alex , terasurfer , Edmond Seymore, Oscar Rangel, Luke Pendergrass, Asp the Wyvern, Junyu Yang, David Flickinger, Luke, Spiking Neurons AB, subjectnull, Pyrater, Nikolai Manek, senxiiz, Ajan Kanaga, Johann-Peter Hartmann, Artur Olbinski, Kevin Schuppel, Derek Yates, Kalila, K, Talal Aujan, Khalefa Al-Ahmad, Gabriel Puliatti, John Villwock, WelcomeToTheClub, Daniel P. Andersen, Preetika Verma, Deep Realms, Fen Risland, trip7s trip, webtim, Sean Connelly, Michael Levine, Chris McCloskey, biorpg, vamX, Viktor Bowallius, Cory Kujawski.
890
+
891
+ Thank you to all my generous patrons and donaters!
892
+
893
+ <!-- footer end -->
894
+
895
+ # Original model card: NousResearch's Redmond Hermes Coder
896
+
897
+
898
+ # Model Card: Redmond-Hermes-Coder 15B
899
+
900
+
901
+ ![xmtf](https://github.com/bigscience-workshop/xmtf/blob/master/xmtf_banner.png?raw=true)
902
+
903
+ # Table of Contents
904
+
905
+ 1. [Model Summary](#model-summary)
906
+ 2. [Use](#use)
907
+ 3. [Limitations](#limitations)
908
+ 4. [Training](#training)
909
+ 5. [Evaluation](#evaluation)
910
+ 7. [Citation](#citation)
911
+
912
+ # Model Summary
913
+
914
+ > We present BLOOMZ & mT0, a family of models capable of following human instructions in dozens of languages zero-shot. We finetune BLOOM & mT5 pretrained multilingual language models on our crosslingual task mixture (xP3) and find the resulting models capable of crosslingual generalization to unseen tasks & languages.
915
+
916
+ - **Repository:** [bigscience-workshop/xmtf](https://github.com/bigscience-workshop/xmtf)
917
+ - **Paper:** [Crosslingual Generalization through Multitask Finetuning](https://arxiv.org/abs/2211.01786)
918
+ - **Point of Contact:** [Niklas Muennighoff](mailto:niklas@hf.co)
919
+ - **Languages:** Refer to [bloom](https://huggingface.co/bigscience/bloom) for pretraining & [xP3](https://huggingface.co/datasets/bigscience/xP3) for finetuning language proportions. It understands both pretraining & finetuning languages.
920
+ - **BLOOMZ & mT0 Model Family:**
921
+
922
+ <div class="max-w-full overflow-auto">
923
+ <table>
924
+ <tr>
925
+ <th colspan="12">Multitask finetuned on <a style="font-weight:bold" href=https://huggingface.co/datasets/bigscience/xP3>xP3</a>. Recommended for prompting in English.
926
+ </tr>
927
+ <tr>
928
+ <td>Parameters</td>
929
+ <td>300M</td>
930
+ <td>580M</td>
931
+ <td>1.2B</td>
932
+ <td>3.7B</td>
933
+ <td>13B</td>
934
+ <td>560M</td>
935
+ <td>1.1B</td>
936
+ <td>1.7B</td>
937
+ <td>3B</td>
938
+ <td>7.1B</td>
939
+ <td>176B</td>
940
+ </tr>
941
+ <tr>
942
+ <td>Finetuned Model</td>
943
+ <td><a href=https://huggingface.co/bigscience/mt0-small>mt0-small</a></td>
944
+ <td><a href=https://huggingface.co/bigscience/mt0-base>mt0-base</a></td>
945
+ <td><a href=https://huggingface.co/bigscience/mt0-large>mt0-large</a></td>
946
+ <td><a href=https://huggingface.co/bigscience/mt0-xl>mt0-xl</a></td>
947
+ <td><a href=https://huggingface.co/bigscience/mt0-xxl>mt0-xxl</a></td>
948
+ <td><a href=https://huggingface.co/bigscience/bloomz-560m>bloomz-560m</a></td>
949
+ <td><a href=https://huggingface.co/bigscience/bloomz-1b1>bloomz-1b1</a></td>
950
+ <td><a href=https://huggingface.co/bigscience/bloomz-1b7>bloomz-1b7</a></td>
951
+ <td><a href=https://huggingface.co/bigscience/bloomz-3b>bloomz-3b</a></td>
952
+ <td><a href=https://huggingface.co/bigscience/bloomz-7b1>bloomz-7b1</a></td>
953
+ <td><a href=https://huggingface.co/bigscience/bloomz>bloomz</a></td>
954
+ </tr>
955
+ </tr>
956
+ <tr>
957
+ <th colspan="12">Multitask finetuned on <a style="font-weight:bold" href=https://huggingface.co/datasets/bigscience/xP3mt>xP3mt</a>. Recommended for prompting in non-English.</th>
958
+ </tr>
959
+ <tr>
960
+ <td>Finetuned Model</td>
961
+ <td></td>
962
+ <td></td>
963
+ <td></td>
964
+ <td></td>
965
+ <td><a href=https://huggingface.co/bigscience/mt0-xxl-mt>mt0-xxl-mt</a></td>
966
+ <td></td>
967
+ <td></td>
968
+ <td></td>
969
+ <td></td>
970
+ <td><a href=https://huggingface.co/bigscience/bloomz-7b1-mt>bloomz-7b1-mt</a></td>
971
+ <td><a href=https://huggingface.co/bigscience/bloomz-mt>bloomz-mt</a></td>
972
+ </tr>
973
+ <th colspan="12">Multitask finetuned on <a style="font-weight:bold" href=https://huggingface.co/datasets/Muennighoff/P3>P3</a>. Released for research purposes only. Strictly inferior to above models!</th>
974
+ </tr>
975
+ <tr>
976
+ <td>Finetuned Model</td>
977
+ <td></td>
978
+ <td></td>
979
+ <td></td>
980
+ <td></td>
981
+ <td><a href=https://huggingface.co/bigscience/mt0-xxl-p3>mt0-xxl-p3</a></td>
982
+ <td></td>
983
+ <td></td>
984
+ <td></td>
985
+ <td></td>
986
+ <td><a href=https://huggingface.co/bigscience/bloomz-7b1-p3>bloomz-7b1-p3</a></td>
987
+ <td><a href=https://huggingface.co/bigscience/bloomz-p3>bloomz-p3</a></td>
988
+ </tr>
989
+ <th colspan="12">Original pretrained checkpoints. Not recommended.</th>
990
+ <tr>
991
+ <td>Pretrained Model</td>
992
+ <td><a href=https://huggingface.co/google/mt5-small>mt5-small</a></td>
993
+ <td><a href=https://huggingface.co/google/mt5-base>mt5-base</a></td>
994
+ <td><a href=https://huggingface.co/google/mt5-large>mt5-large</a></td>
995
+ <td><a href=https://huggingface.co/google/mt5-xl>mt5-xl</a></td>
996
+ <td><a href=https://huggingface.co/google/mt5-xxl>mt5-xxl</a></td>
997
+ <td><a href=https://huggingface.co/bigscience/bloom-560m>bloom-560m</a></td>
998
+ <td><a href=https://huggingface.co/bigscience/bloom-1b1>bloom-1b1</a></td>
999
+ <td><a href=https://huggingface.co/bigscience/bloom-1b7>bloom-1b7</a></td>
1000
+ <td><a href=https://huggingface.co/bigscience/bloom-3b>bloom-3b</a></td>
1001
+ <td><a href=https://huggingface.co/bigscience/bloom-7b1>bloom-7b1</a></td>
1002
+ <td><a href=https://huggingface.co/bigscience/bloom>bloom</a></td>
1003
+ </tr>
1004
+ </table>
1005
+ </div>
1006
+
1007
+
1008
+ # Use
1009
+
1010
+ ## Intended use
1011
+
1012
+ We recommend using the model to perform tasks expressed in natural language. For example, given the prompt "*Translate to English: Je t’aime.*", the model will most likely answer "*I love you.*". Some prompt ideas from our paper:
1013
+ - 一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?
1014
+ - Suggest at least five related search terms to "Mạng neural nhân tạo".
1015
+ - Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is "Heroes Come in All Shapes and Sizes". Story (in Spanish):
1016
+ - Explain in a sentence in Telugu what is backpropagation in neural networks.
1017
+
1018
+ **Feel free to share your generations in the Community tab!**
1019
+
1020
+ ## How to use
1021
+
1022
+ ### CPU
1023
+
1024
+ <details>
1025
+ <summary> Click to expand </summary>
1026
+
1027
+ ```python
1028
+ # pip install -q transformers
1029
+ from transformers import AutoModelForCausalLM, AutoTokenizer
1030
+
1031
+ checkpoint = "bigscience/bloomz"
1032
+
1033
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint)
1034
+ model = AutoModelForCausalLM.from_pretrained(checkpoint)
1035
+
1036
+ inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt")
1037
+ outputs = model.generate(inputs)
1038
+ print(tokenizer.decode(outputs[0]))
1039
+ ```
1040
+
1041
+ </details>
1042
+
1043
+ ### GPU
1044
+
1045
+ <details>
1046
+ <summary> Click to expand </summary>
1047
+
1048
+ ```python
1049
+ # pip install -q transformers accelerate
1050
+ from transformers import AutoModelForCausalLM, AutoTokenizer
1051
+
1052
+ checkpoint = "bigscience/bloomz"
1053
+
1054
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint)
1055
+ model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype="auto", device_map="auto")
1056
+
1057
+ inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
1058
+ outputs = model.generate(inputs)
1059
+ print(tokenizer.decode(outputs[0]))
1060
+ ```
1061
+
1062
+ </details>
1063
+
1064
+ ### GPU in 8bit
1065
+
1066
+ <details>
1067
+ <summary> Click to expand </summary>
1068
+
1069
+ ```python
1070
+ # pip install -q transformers accelerate bitsandbytes
1071
+ from transformers import AutoModelForCausalLM, AutoTokenizer
1072
+
1073
+ checkpoint = "bigscience/bloomz"
1074
+
1075
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint)
1076
+ model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", load_in_8bit=True)
1077
+
1078
+ inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
1079
+ outputs = model.generate(inputs)
1080
+ print(tokenizer.decode(outputs[0]))
1081
+ ```
1082
+
1083
+ </details>
1084
+
1085
+ <!-- Necessary for whitespace -->
1086
+ ###
1087
+
1088
+ # Limitations
1089
+
1090
+ **Prompt Engineering:** The performance may vary depending on the prompt. For BLOOMZ models, we recommend making it very clear when the input stops to avoid the model trying to continue it. For example, the prompt "*Translate to English: Je t'aime*" without the full stop (.) at the end, may result in the model trying to continue the French sentence. Better prompts are e.g. "*Translate to English: Je t'aime.*", "*Translate to English: Je t'aime. Translation:*" "*What is "Je t'aime." in English?*", where it is clear for the model when it should answer. Further, we recommend providing the model as much context as possible. For example, if you want it to answer in Telugu, then tell the model, e.g. "*Explain in a sentence in Telugu what is backpropagation in neural networks.*".
1091
+
1092
+ # Training
1093
+
1094
+ ## Model
1095
+
1096
+ - **Architecture:** Same as [bloom](https://huggingface.co/bigscience/bloom), also refer to the `config.json` file
1097
+ - **Finetuning steps:** 498
1098
+ - **Finetuning tokens:** 2.09 billion
1099
+ - **Finetuning layout:** 72x pipeline parallel, 1x tensor parallel, 4x data parallel
1100
+ - **Precision:** bfloat16
1101
+
1102
+ ## Hardware
1103
+
1104
+ - **CPUs:** AMD CPUs with 512GB memory per node
1105
+ - **GPUs:** 288 A100 80GB GPUs with 8 GPUs per node (36 nodes) using NVLink 4 inter-gpu connects, 4 OmniPath links
1106
+ - **Communication:** NCCL-communications network with a fully dedicated subnet
1107
+
1108
+ ## Software
1109
+
1110
+ - **Orchestration:** [Megatron-DeepSpeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed)
1111
+ - **Optimizer & parallelism:** [DeepSpeed](https://github.com/microsoft/DeepSpeed)
1112
+ - **Neural networks:** [PyTorch](https://github.com/pytorch/pytorch) (pytorch-1.11 w/ CUDA-11.5)
1113
+ - **FP16 if applicable:** [apex](https://github.com/NVIDIA/apex)
1114
+
1115
+ # Evaluation
1116
+
1117
+ We refer to Table 7 from our [paper](https://arxiv.org/abs/2211.01786) & [bigscience/evaluation-results](https://huggingface.co/datasets/bigscience/evaluation-results) for zero-shot results on unseen tasks. The sidebar reports zero-shot performance of the best prompt per dataset config.
1118
+
1119
+ # Citation
1120
+ ```bibtex
1121
+ @article{muennighoff2022crosslingual,
1122
+ title={Crosslingual generalization through multitask finetuning},
1123
+ author={Muennighoff, Niklas and Wang, Thomas and Sutawika, Lintang and Roberts, Adam and Biderman, Stella and Scao, Teven Le and Bari, M Saiful and Shen, Sheng and Yong, Zheng-Xin and Schoelkopf, Hailey and others},
1124
+ journal={arXiv preprint arXiv:2211.01786},
1125
+ year={2022}
1126
+ }
1127
+ ```