TheBloke commited on
Commit
ff198bd
1 Parent(s): 491a53d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +365 -48
README.md CHANGED
@@ -1,6 +1,256 @@
1
  ---
 
2
  inference: false
3
- license: other
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  ---
5
 
6
  <!-- header start -->
@@ -21,12 +271,7 @@ license: other
21
 
22
  These files are GGML format model files for [Bigcode's Starcoder](https://huggingface.co/bigcode/starcoder).
23
 
24
- GGML files are for CPU + GPU inference using [llama.cpp](https://github.com/ggerganov/llama.cpp) and libraries and UIs which support this format, such as:
25
- * [text-generation-webui](https://github.com/oobabooga/text-generation-webui)
26
- * [KoboldCpp](https://github.com/LostRuins/koboldcpp)
27
- * [ParisNeo/GPT4All-UI](https://github.com/ParisNeo/gpt4all-ui)
28
- * [llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
29
- * [ctransformers](https://github.com/marella/ctransformers)
30
 
31
  ## Repositories available
32
 
@@ -35,31 +280,23 @@ GGML files are for CPU + GPU inference using [llama.cpp](https://github.com/gger
35
  * [Unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/bigcode/starcoder)
36
 
37
  <!-- compatibility_ggml start -->
38
- ## Compatibility
39
 
40
- ### Original llama.cpp quant methods: `q4_0, q4_1, q5_0, q5_1, q8_0`
41
 
42
- I have quantized these 'original' quantisation methods using an older version of llama.cpp so that they remain compatible with llama.cpp as of May 19th, commit `2d5db48`.
 
 
 
 
 
43
 
44
- They should be compatible with all current UIs and libraries that use llama.cpp, such as those listed at the top of this README.
45
 
46
- ### New k-quant methods: `q2_K, q3_K_S, q3_K_M, q3_K_L, q4_K_S, q4_K_M, q5_K_S, q6_K`
47
 
48
- These new quantisation methods are only compatible with llama.cpp as of June 6th, commit `2d43387`.
49
-
50
- They will NOT be compatible with koboldcpp, text-generation-ui, and other UIs and libraries yet. Support is expected to come over the next few days.
51
-
52
- ## Explanation of the new k-quant methods
53
-
54
- The new methods available are:
55
- * GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
56
- * GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
57
- * GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
58
- * GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
59
- * GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
60
- * GGML_TYPE_Q8_K - "type-0" 8-bit quantization. Only used for quantizing intermediate results. The difference to the existing Q8_0 is that the block size is 256. All 2-6 bit dot products are implemented for this quantization type.
61
-
62
- Refer to the Provided Files table below to see what files use which methods, and how.
63
  <!-- compatibility_ggml end -->
64
 
65
  ## Provided files
@@ -71,26 +308,6 @@ Refer to the Provided Files table below to see what files use which methods, and
71
  | starcoder.ggmlv3.q5_1.bin | q5_1 | 5 | 14.26 GB | 16.76 GB | Original llama.cpp quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
72
  | starcoder.ggmlv3.q8_0.bin | q8_0 | 8 | 20.11 GB | 22.61 GB | Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
73
 
74
-
75
- **Note**: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.
76
-
77
- ## How to run in `llama.cpp`
78
-
79
- I use the following command line; adjust for your tastes and needs:
80
-
81
- ```
82
- ./main -t 10 -ngl 32 -m starcode.ggmlv3.q5_0.bin --color -c 2048 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "### Instruction: Write a story about llamas\n### Response:"
83
- ```
84
- Change `-t 10` to the number of physical CPU cores you have. For example if your system has 8 cores/16 threads, use `-t 8`.
85
-
86
- Change `-ngl 32` to the number of layers to offload to GPU. Remove it if you don't have GPU acceleration.
87
-
88
- If you want to have a chat-style conversation, replace the `-p <PROMPT>` argument with `-i -ins`
89
-
90
- ## How to run in `text-generation-webui`
91
-
92
- Further instructions here: [text-generation-webui/docs/llama.cpp-models.md](https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp-models.md).
93
-
94
  <!-- footer start -->
95
  ## Discord
96
 
@@ -121,4 +338,104 @@ Thank you to all my generous patrons and donaters!
121
 
122
  # Original model card: Bigcode's Starcoder
123
 
124
- No original model card was provided.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ pipeline_tag: text-generation
3
  inference: false
4
+ license: bigcode-openrail-m
5
+ datasets:
6
+ - bigcode/the-stack-dedup
7
+ metrics:
8
+ - code_eval
9
+ library_name: transformers
10
+ tags:
11
+ - code
12
+ model-index:
13
+ - name: StarCoder
14
+ results:
15
+ - task:
16
+ type: text-generation
17
+ dataset:
18
+ type: openai_humaneval
19
+ name: HumanEval (Prompted)
20
+ metrics:
21
+ - name: pass@1
22
+ type: pass@1
23
+ value: 0.408
24
+ verified: false
25
+ - task:
26
+ type: text-generation
27
+ dataset:
28
+ type: openai_humaneval
29
+ name: HumanEval
30
+ metrics:
31
+ - name: pass@1
32
+ type: pass@1
33
+ value: 0.336
34
+ verified: false
35
+ - task:
36
+ type: text-generation
37
+ dataset:
38
+ type: mbpp
39
+ name: MBPP
40
+ metrics:
41
+ - name: pass@1
42
+ type: pass@1
43
+ value: 0.527
44
+ verified: false
45
+ - task:
46
+ type: text-generation
47
+ dataset:
48
+ type: ds1000
49
+ name: DS-1000 (Overall Completion)
50
+ metrics:
51
+ - name: pass@1
52
+ type: pass@1
53
+ value: 0.26
54
+ verified: false
55
+ - task:
56
+ type: text-generation
57
+ dataset:
58
+ type: nuprl/MultiPL-E
59
+ name: MultiPL-HumanEval (C++)
60
+ metrics:
61
+ - name: pass@1
62
+ type: pass@1
63
+ value: 0.3155
64
+ verified: false
65
+ - task:
66
+ type: text-generation
67
+ dataset:
68
+ type: nuprl/MultiPL-E
69
+ name: MultiPL-HumanEval (C#)
70
+ metrics:
71
+ - name: pass@1
72
+ type: pass@1
73
+ value: 0.2101
74
+ verified: false
75
+ - task:
76
+ type: text-generation
77
+ dataset:
78
+ type: nuprl/MultiPL-E
79
+ name: MultiPL-HumanEval (D)
80
+ metrics:
81
+ - name: pass@1
82
+ type: pass@1
83
+ value: 0.1357
84
+ verified: false
85
+ - task:
86
+ type: text-generation
87
+ dataset:
88
+ type: nuprl/MultiPL-E
89
+ name: MultiPL-HumanEval (Go)
90
+ metrics:
91
+ - name: pass@1
92
+ type: pass@1
93
+ value: 0.1761
94
+ verified: false
95
+ - task:
96
+ type: text-generation
97
+ dataset:
98
+ type: nuprl/MultiPL-E
99
+ name: MultiPL-HumanEval (Java)
100
+ metrics:
101
+ - name: pass@1
102
+ type: pass@1
103
+ value: 0.3022
104
+ verified: false
105
+ - task:
106
+ type: text-generation
107
+ dataset:
108
+ type: nuprl/MultiPL-E
109
+ name: MultiPL-HumanEval (Julia)
110
+ metrics:
111
+ - name: pass@1
112
+ type: pass@1
113
+ value: 0.2302
114
+ verified: false
115
+ - task:
116
+ type: text-generation
117
+ dataset:
118
+ type: nuprl/MultiPL-E
119
+ name: MultiPL-HumanEval (JavaScript)
120
+ metrics:
121
+ - name: pass@1
122
+ type: pass@1
123
+ value: 0.3079
124
+ verified: false
125
+ - task:
126
+ type: text-generation
127
+ dataset:
128
+ type: nuprl/MultiPL-E
129
+ name: MultiPL-HumanEval (Lua)
130
+ metrics:
131
+ - name: pass@1
132
+ type: pass@1
133
+ value: 0.2389
134
+ verified: false
135
+ - task:
136
+ type: text-generation
137
+ dataset:
138
+ type: nuprl/MultiPL-E
139
+ name: MultiPL-HumanEval (PHP)
140
+ metrics:
141
+ - name: pass@1
142
+ type: pass@1
143
+ value: 0.2608
144
+ verified: false
145
+ - task:
146
+ type: text-generation
147
+ dataset:
148
+ type: nuprl/MultiPL-E
149
+ name: MultiPL-HumanEval (Perl)
150
+ metrics:
151
+ - name: pass@1
152
+ type: pass@1
153
+ value: 0.1734
154
+ verified: false
155
+ - task:
156
+ type: text-generation
157
+ dataset:
158
+ type: nuprl/MultiPL-E
159
+ name: MultiPL-HumanEval (Python)
160
+ metrics:
161
+ - name: pass@1
162
+ type: pass@1
163
+ value: 0.3357
164
+ verified: false
165
+ - task:
166
+ type: text-generation
167
+ dataset:
168
+ type: nuprl/MultiPL-E
169
+ name: MultiPL-HumanEval (R)
170
+ metrics:
171
+ - name: pass@1
172
+ type: pass@1
173
+ value: 0.155
174
+ verified: false
175
+ - task:
176
+ type: text-generation
177
+ dataset:
178
+ type: nuprl/MultiPL-E
179
+ name: MultiPL-HumanEval (Ruby)
180
+ metrics:
181
+ - name: pass@1
182
+ type: pass@1
183
+ value: 0.0124
184
+ verified: false
185
+ - task:
186
+ type: text-generation
187
+ dataset:
188
+ type: nuprl/MultiPL-E
189
+ name: MultiPL-HumanEval (Racket)
190
+ metrics:
191
+ - name: pass@1
192
+ type: pass@1
193
+ value: 0.0007
194
+ verified: false
195
+ - task:
196
+ type: text-generation
197
+ dataset:
198
+ type: nuprl/MultiPL-E
199
+ name: MultiPL-HumanEval (Rust)
200
+ metrics:
201
+ - name: pass@1
202
+ type: pass@1
203
+ value: 0.2184
204
+ verified: false
205
+ - task:
206
+ type: text-generation
207
+ dataset:
208
+ type: nuprl/MultiPL-E
209
+ name: MultiPL-HumanEval (Scala)
210
+ metrics:
211
+ - name: pass@1
212
+ type: pass@1
213
+ value: 0.2761
214
+ verified: false
215
+ - task:
216
+ type: text-generation
217
+ dataset:
218
+ type: nuprl/MultiPL-E
219
+ name: MultiPL-HumanEval (Bash)
220
+ metrics:
221
+ - name: pass@1
222
+ type: pass@1
223
+ value: 0.1046
224
+ verified: false
225
+ - task:
226
+ type: text-generation
227
+ dataset:
228
+ type: nuprl/MultiPL-E
229
+ name: MultiPL-HumanEval (Swift)
230
+ metrics:
231
+ - name: pass@1
232
+ type: pass@1
233
+ value: 0.2274
234
+ verified: false
235
+ - task:
236
+ type: text-generation
237
+ dataset:
238
+ type: nuprl/MultiPL-E
239
+ name: MultiPL-HumanEval (TypeScript)
240
+ metrics:
241
+ - name: pass@1
242
+ type: pass@1
243
+ value: 0.3229
244
+ verified: false
245
+ extra_gated_prompt: >-
246
+ ## Model License Agreement
247
+
248
+ Please read the BigCode [OpenRAIL-M
249
+ license](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement)
250
+ agreement before accepting it.
251
+
252
+ extra_gated_fields:
253
+ I accept the above license agreement, and will use the Model complying with the set of use restrictions and sharing requirements: checkbox
254
  ---
255
 
256
  <!-- header start -->
 
271
 
272
  These files are GGML format model files for [Bigcode's Starcoder](https://huggingface.co/bigcode/starcoder).
273
 
274
+ Please note that these GGMLs are **not compatbile with llama.cpp**. Please see below for a list of tools known to work with these model files.
 
 
 
 
 
275
 
276
  ## Repositories available
277
 
 
280
  * [Unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/bigcode/starcoder)
281
 
282
  <!-- compatibility_ggml start -->
283
+ ## Compatibilty
284
 
285
+ These files are **not** compatible with llama.cpp.
286
 
287
+ Currently they can be used with:
288
+ * KoboldCpp, a powerful inference engine based on llama.cpp, with good UI: [KoboldCpp](https://github.com/LostRuins/koboldcpp)
289
+ * The ctransformers Python library, which includes LangChain support: [ctransformers](https://github.com/marella/ctransformers)
290
+ * The GPT4All-UI which uses ctransformers: [GPT4All-UI](https://github.com/ParisNeo/gpt4all-ui)
291
+ * [rustformers' llm](https://github.com/rustformers/llm)
292
+ * The example `mpt` binary provided with [ggml](https://github.com/ggerganov/ggml)
293
 
294
+ As other options become available I will endeavour to update them here (do let me know in the Community tab if I've missed something!)
295
 
296
+ ## Tutorial for using GPT4All-UI
297
 
298
+ * [Text tutorial, written by **Lucas3DCG**](https://huggingface.co/TheBloke/MPT-7B-Storywriter-GGML/discussions/2#6475d914e9b57ce0caa68888)
299
+ * [Video tutorial, by GPT4All-UI's author **ParisNeo**](https://www.youtube.com/watch?v=ds_U0TDzbzI)
 
 
 
 
 
 
 
 
 
 
 
 
 
300
  <!-- compatibility_ggml end -->
301
 
302
  ## Provided files
 
308
  | starcoder.ggmlv3.q5_1.bin | q5_1 | 5 | 14.26 GB | 16.76 GB | Original llama.cpp quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
309
  | starcoder.ggmlv3.q8_0.bin | q8_0 | 8 | 20.11 GB | 22.61 GB | Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
310
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
311
  <!-- footer start -->
312
  ## Discord
313
 
 
338
 
339
  # Original model card: Bigcode's Starcoder
340
 
341
+ # StarCoder
342
+
343
+ ![banner](https://huggingface.co/datasets/bigcode/admin/resolve/main/StarCoderBanner.png)
344
+
345
+ Play with the model on the [StarCoder Playground](https://huggingface.co/spaces/bigcode/bigcode-playground).
346
+
347
+ ## Table of Contents
348
+
349
+ 1. [Model Summary](##model-summary)
350
+ 2. [Use](##use)
351
+ 3. [Limitations](##limitations)
352
+ 4. [Training](##training)
353
+ 5. [License](##license)
354
+ 6. [Citation](##citation)
355
+
356
+ ## Model Summary
357
+
358
+ The StarCoder models are 15.5B parameter models trained on 80+ programming languages from [The Stack (v1.2)](https://huggingface.co/datasets/bigcode/the-stack), with opt-out requests excluded. The model uses [Multi Query Attention](https://arxiv.org/abs/1911.02150), [a context window of 8192 tokens](https://arxiv.org/abs/2205.14135), and was trained using the [Fill-in-the-Middle objective](https://arxiv.org/abs/2207.14255) on 1 trillion tokens.
359
+
360
+ - **Repository:** [bigcode/Megatron-LM](https://github.com/bigcode-project/Megatron-LM)
361
+ - **Project Website:** [bigcode-project.org](https://www.bigcode-project.org)
362
+ - **Paper:** [💫StarCoder: May the source be with you!](https://arxiv.org/abs/2305.06161)
363
+ - **Point of Contact:** [contact@bigcode-project.org](mailto:contact@bigcode-project.org)
364
+ - **Languages:** 80+ Programming languages
365
+
366
+
367
+ ## Use
368
+
369
+ ### Intended use
370
+
371
+ The model was trained on GitHub code. As such it is _not_ an instruction model and commands like "Write a function that computes the square root." do not work well. However, by using the [Tech Assistant prompt](https://huggingface.co/datasets/bigcode/ta-prompt) you can turn it into a capable technical assistant.
372
+
373
+ **Feel free to share your generations in the Community tab!**
374
+
375
+ ### Generation
376
+ ```python
377
+ # pip install -q transformers
378
+ from transformers import AutoModelForCausalLM, AutoTokenizer
379
+
380
+ checkpoint = "bigcode/starcoder"
381
+ device = "cuda" # for GPU usage or "cpu" for CPU usage
382
+
383
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint)
384
+ model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
385
+
386
+ inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
387
+ outputs = model.generate(inputs)
388
+ print(tokenizer.decode(outputs[0]))
389
+ ```
390
+
391
+ ### Fill-in-the-middle
392
+ Fill-in-the-middle uses special tokens to identify the prefix/middle/suffix part of the input and output:
393
+
394
+ ```python
395
+ input_text = "<fim_prefix>def print_hello_world():\n <fim_suffix>\n print('Hello world!')<fim_middle>"
396
+ inputs = tokenizer.encode(input_text, return_tensors="pt").to(device)
397
+ outputs = model.generate(inputs)
398
+ print(tokenizer.decode(outputs[0]))
399
+ ```
400
+
401
+ ### Attribution & Other Requirements
402
+
403
+ The pretraining dataset of the model was filtered for permissive licenses only. Nevertheless, the model can generate source code verbatim from the dataset. The code's license might require attribution and/or other specific requirements that must be respected. We provide a [search index](https://huggingface.co/spaces/bigcode/starcoder-search) that let's you search through the pretraining data to identify where generated code came from and apply the proper attribution to your code.
404
+
405
+ # Limitations
406
+
407
+ The model has been trained on source code from 80+ programming languages. The predominant natural language in source code is English although other languages are also present. As such the model is capable of generating code snippets provided some context but the generated code is not guaranteed to work as intended. It can be inefficient, contain bugs or exploits. See [the paper](https://drive.google.com/file/d/1cN-b9GnWtHzQRoE7M7gAEyivY0kl4BYs/view) for an in-depth discussion of the model limitations.
408
+
409
+ # Training
410
+
411
+ ## Model
412
+
413
+ - **Architecture:** GPT-2 model with multi-query attention and Fill-in-the-Middle objective
414
+ - **Pretraining steps:** 250k
415
+ - **Pretraining tokens:** 1 trillion
416
+ - **Precision:** bfloat16
417
+
418
+ ## Hardware
419
+
420
+ - **GPUs:** 512 Tesla A100
421
+ - **Training time:** 24 days
422
+
423
+ ## Software
424
+
425
+ - **Orchestration:** [Megatron-LM](https://github.com/bigcode-project/Megatron-LM)
426
+ - **Neural networks:** [PyTorch](https://github.com/pytorch/pytorch)
427
+ - **BP16 if applicable:** [apex](https://github.com/NVIDIA/apex)
428
+
429
+ # License
430
+ The model is licensed under the BigCode OpenRAIL-M v1 license agreement. You can find the full agreement [here](https://huggingface.co/spaces/bigcode/bigcode-model-license-agreement).
431
+ # Citation
432
+ ```
433
+ @article{li2023starcoder,
434
+ title={StarCoder: may the source be with you!},
435
+ author={Raymond Li and Loubna Ben Allal and Yangtian Zi and Niklas Muennighoff and Denis Kocetkov and Chenghao Mou and Marc Marone and Christopher Akiki and Jia Li and Jenny Chim and Qian Liu and Evgenii Zheltonozhskii and Terry Yue Zhuo and Thomas Wang and Olivier Dehaene and Mishig Davaadorj and Joel Lamy-Poirier and João Monteiro and Oleh Shliazhko and Nicolas Gontier and Nicholas Meade and Armel Zebaze and Ming-Ho Yee and Logesh Kumar Umapathi and Jian Zhu and Benjamin Lipkin and Muhtasham Oblokulov and Zhiruo Wang and Rudra Murthy and Jason Stillerman and Siva Sankalp Patel and Dmitry Abulkhanov and Marco Zocca and Manan Dey and Zhihan Zhang and Nour Fahmy and Urvashi Bhattacharyya and Wenhao Yu and Swayam Singh and Sasha Luccioni and Paulo Villegas and Maxim Kunakov and Fedor Zhdanov and Manuel Romero and Tony Lee and Nadav Timor and Jennifer Ding and Claire Schlesinger and Hailey Schoelkopf and Jan Ebert and Tri Dao and Mayank Mishra and Alex Gu and Jennifer Robinson and Carolyn Jane Anderson and Brendan Dolan-Gavitt and Danish Contractor and Siva Reddy and Daniel Fried and Dzmitry Bahdanau and Yacine Jernite and Carlos Muñoz Ferrandis and Sean Hughes and Thomas Wolf and Arjun Guha and Leandro von Werra and Harm de Vries},
436
+ year={2023},
437
+ eprint={2305.06161},
438
+ archivePrefix={arXiv},
439
+ primaryClass={cs.CL}
440
+ }
441
+ ```