Commit
β’
d61ee7d
1
Parent(s):
aa6bc71
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,428 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
license: other
|
5 |
+
tags:
|
6 |
+
- axolotl
|
7 |
+
- generated_from_trainer
|
8 |
+
- instruct
|
9 |
+
- finetune
|
10 |
+
- chatml
|
11 |
+
- gpt4
|
12 |
+
- synthetic data
|
13 |
+
- science
|
14 |
+
- physics
|
15 |
+
- chemistry
|
16 |
+
- biology
|
17 |
+
- math
|
18 |
+
- llama
|
19 |
+
- llama3
|
20 |
+
base_model: Weyaxi/Einstein-v6.1-Llama3-8B
|
21 |
+
datasets:
|
22 |
+
- allenai/ai2_arc
|
23 |
+
- camel-ai/physics
|
24 |
+
- camel-ai/chemistry
|
25 |
+
- camel-ai/biology
|
26 |
+
- camel-ai/math
|
27 |
+
- metaeval/reclor
|
28 |
+
- openbookqa
|
29 |
+
- mandyyyyii/scibench
|
30 |
+
- derek-thomas/ScienceQA
|
31 |
+
- TIGER-Lab/ScienceEval
|
32 |
+
- jondurbin/airoboros-3.2
|
33 |
+
- LDJnr/Capybara
|
34 |
+
- Cot-Alpaca-GPT4-From-OpenHermes-2.5
|
35 |
+
- STEM-AI-mtl/Electrical-engineering
|
36 |
+
- knowrohit07/saraswati-stem
|
37 |
+
- sablo/oasst2_curated
|
38 |
+
- lmsys/lmsys-chat-1m
|
39 |
+
- TIGER-Lab/MathInstruct
|
40 |
+
- bigbio/med_qa
|
41 |
+
- meta-math/MetaMathQA-40K
|
42 |
+
- openbookqa
|
43 |
+
- piqa
|
44 |
+
- metaeval/reclor
|
45 |
+
- derek-thomas/ScienceQA
|
46 |
+
- scibench
|
47 |
+
- sciq
|
48 |
+
- Open-Orca/SlimOrca
|
49 |
+
- migtissera/Synthia-v1.3
|
50 |
+
- TIGER-Lab/ScienceEval
|
51 |
+
- allenai/WildChat
|
52 |
+
- microsoft/orca-math-word-problems-200k
|
53 |
+
- openchat/openchat_sharegpt4_dataset
|
54 |
+
- teknium/GPTeacher-General-Instruct
|
55 |
+
- m-a-p/CodeFeedback-Filtered-Instruction
|
56 |
+
- totally-not-an-llm/EverythingLM-data-V3
|
57 |
+
- HuggingFaceH4/no_robots
|
58 |
+
- OpenAssistant/oasst_top1_2023-08-25
|
59 |
+
- WizardLM/WizardLM_evol_instruct_70k
|
60 |
+
model-index:
|
61 |
+
- name: Einstein-v6.1-Llama3-8B
|
62 |
+
results:
|
63 |
+
- task:
|
64 |
+
type: text-generation
|
65 |
+
name: Text Generation
|
66 |
+
dataset:
|
67 |
+
name: AI2 Reasoning Challenge (25-Shot)
|
68 |
+
type: ai2_arc
|
69 |
+
config: ARC-Challenge
|
70 |
+
split: test
|
71 |
+
args:
|
72 |
+
num_few_shot: 25
|
73 |
+
metrics:
|
74 |
+
- type: acc_norm
|
75 |
+
value: 62.46
|
76 |
+
name: normalized accuracy
|
77 |
+
source:
|
78 |
+
url: >-
|
79 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
80 |
+
name: Open LLM Leaderboard
|
81 |
+
- task:
|
82 |
+
type: text-generation
|
83 |
+
name: Text Generation
|
84 |
+
dataset:
|
85 |
+
name: HellaSwag (10-Shot)
|
86 |
+
type: hellaswag
|
87 |
+
split: validation
|
88 |
+
args:
|
89 |
+
num_few_shot: 10
|
90 |
+
metrics:
|
91 |
+
- type: acc_norm
|
92 |
+
value: 82.41
|
93 |
+
name: normalized accuracy
|
94 |
+
source:
|
95 |
+
url: >-
|
96 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
97 |
+
name: Open LLM Leaderboard
|
98 |
+
- task:
|
99 |
+
type: text-generation
|
100 |
+
name: Text Generation
|
101 |
+
dataset:
|
102 |
+
name: MMLU (5-Shot)
|
103 |
+
type: cais/mmlu
|
104 |
+
config: all
|
105 |
+
split: test
|
106 |
+
args:
|
107 |
+
num_few_shot: 5
|
108 |
+
metrics:
|
109 |
+
- type: acc
|
110 |
+
value: 66.19
|
111 |
+
name: accuracy
|
112 |
+
source:
|
113 |
+
url: >-
|
114 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
115 |
+
name: Open LLM Leaderboard
|
116 |
+
- task:
|
117 |
+
type: text-generation
|
118 |
+
name: Text Generation
|
119 |
+
dataset:
|
120 |
+
name: TruthfulQA (0-shot)
|
121 |
+
type: truthful_qa
|
122 |
+
config: multiple_choice
|
123 |
+
split: validation
|
124 |
+
args:
|
125 |
+
num_few_shot: 0
|
126 |
+
metrics:
|
127 |
+
- type: mc2
|
128 |
+
value: 55.1
|
129 |
+
source:
|
130 |
+
url: >-
|
131 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
132 |
+
name: Open LLM Leaderboard
|
133 |
+
- task:
|
134 |
+
type: text-generation
|
135 |
+
name: Text Generation
|
136 |
+
dataset:
|
137 |
+
name: Winogrande (5-shot)
|
138 |
+
type: winogrande
|
139 |
+
config: winogrande_xl
|
140 |
+
split: validation
|
141 |
+
args:
|
142 |
+
num_few_shot: 5
|
143 |
+
metrics:
|
144 |
+
- type: acc
|
145 |
+
value: 79.32
|
146 |
+
name: accuracy
|
147 |
+
source:
|
148 |
+
url: >-
|
149 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
150 |
+
name: Open LLM Leaderboard
|
151 |
+
- task:
|
152 |
+
type: text-generation
|
153 |
+
name: Text Generation
|
154 |
+
dataset:
|
155 |
+
name: GSM8k (5-shot)
|
156 |
+
type: gsm8k
|
157 |
+
config: main
|
158 |
+
split: test
|
159 |
+
args:
|
160 |
+
num_few_shot: 5
|
161 |
+
metrics:
|
162 |
+
- type: acc
|
163 |
+
value: 66.11
|
164 |
+
name: accuracy
|
165 |
+
source:
|
166 |
+
url: >-
|
167 |
+
https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Weyaxi/Einstein-v6.1-Llama3-8B
|
168 |
+
name: Open LLM Leaderboard
|
169 |
+
pipeline_tag: text-generation
|
170 |
+
---
|
171 |
+
|
172 |
+
# π¬ Einstein-v6.1-Llama3-8B - GGUF
|
173 |
+
|
174 |
+
- This is quantized version of [Weyaxi/Einstein-v6.1-Llama3-8B](https://huggingface.co/Weyaxi/Einstein-v6.1-Llama3-8B) created using llama.cpp
|
175 |
+
|
176 |
+
# Model Description
|
177 |
+
|
178 |
+
This model is a full fine-tuned version of [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) on diverse datasets.
|
179 |
+
|
180 |
+
This model is finetuned using `8xRTX3090` + `1xRTXA6000` using [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl).
|
181 |
+
|
182 |
+
This model's training was sponsored by [sablo.ai](https://sablo.ai).
|
183 |
+
|
184 |
+
<details><summary>See axolotl config</summary>
|
185 |
+
|
186 |
+
axolotl version: `0.4.0`
|
187 |
+
```yaml
|
188 |
+
base_model: meta-llama/Meta-Llama-3-8B
|
189 |
+
model_type: LlamaForCausalLM
|
190 |
+
tokenizer_type: AutoTokenizer
|
191 |
+
|
192 |
+
load_in_8bit: false
|
193 |
+
load_in_4bit: false
|
194 |
+
strict: false
|
195 |
+
|
196 |
+
chat_template: chatml
|
197 |
+
datasets:
|
198 |
+
- path: data/merged_all.json
|
199 |
+
ds_type: json
|
200 |
+
type: alpaca
|
201 |
+
conversation: chatml
|
202 |
+
|
203 |
+
- path: data/gpteacher-instruct-special-alpaca.json
|
204 |
+
ds_type: json
|
205 |
+
type: gpteacher
|
206 |
+
conversation: chatml
|
207 |
+
|
208 |
+
- path: data/wizardlm_evol_instruct_70k_random_half.json
|
209 |
+
ds_type: json
|
210 |
+
type: alpaca
|
211 |
+
conversation: chatml
|
212 |
+
|
213 |
+
- path: data/capybara_sharegpt.json
|
214 |
+
ds_type: json
|
215 |
+
type: sharegpt
|
216 |
+
conversation: chatml
|
217 |
+
|
218 |
+
- path: data/synthia-v1.3_sharegpt_12500.json
|
219 |
+
ds_type: json
|
220 |
+
type: sharegpt
|
221 |
+
conversation: chatml
|
222 |
+
|
223 |
+
- path: data/cot_alpaca_gpt4_extracted_openhermes_2.5_sharegpt.json
|
224 |
+
ds_type: json
|
225 |
+
type: sharegpt
|
226 |
+
conversation: chatml
|
227 |
+
|
228 |
+
- path: data/slimorca_dedup_filtered_95k_sharegpt.json
|
229 |
+
ds_type: json
|
230 |
+
type: sharegpt
|
231 |
+
conversation: chatml
|
232 |
+
|
233 |
+
- path: data/airoboros_3.2_without_contextual_slimorca_orca_sharegpt.json
|
234 |
+
ds_type: json
|
235 |
+
type: sharegpt
|
236 |
+
conversation: chatml
|
237 |
+
|
238 |
+
- path: data/allenai_wild_chat_gpt4_english_toxic_random_half_4k_sharegpt.json
|
239 |
+
ds_type: json
|
240 |
+
type: sharegpt
|
241 |
+
strict: false
|
242 |
+
conversation: chatml
|
243 |
+
|
244 |
+
- path: data/pippa_bagel_repo_3k_sharegpt.json
|
245 |
+
ds_type: json
|
246 |
+
type: sharegpt
|
247 |
+
conversation: chatml
|
248 |
+
|
249 |
+
- path: data/gpt4_data_lmys_1m_sharegpt.json
|
250 |
+
ds_type: json
|
251 |
+
type: sharegpt
|
252 |
+
conversation: chatml
|
253 |
+
|
254 |
+
- path: data/sharegpt_gpt4_english.json
|
255 |
+
ds_type: json
|
256 |
+
type: sharegpt
|
257 |
+
conversation: chatml
|
258 |
+
|
259 |
+
- path: data/no_robots_sharegpt.json
|
260 |
+
ds_type: json
|
261 |
+
type: sharegpt
|
262 |
+
strict: false
|
263 |
+
conversation: chatml
|
264 |
+
|
265 |
+
- path: data/oasst_top1_from_fusechatmixture_sharegpt.json
|
266 |
+
ds_type: json
|
267 |
+
type: sharegpt
|
268 |
+
strict: false
|
269 |
+
conversation: chatml
|
270 |
+
|
271 |
+
- path: data/everythinglm-data-v3_sharegpt.json
|
272 |
+
ds_type: json
|
273 |
+
type: sharegpt
|
274 |
+
strict: false
|
275 |
+
conversation: chatml
|
276 |
+
|
277 |
+
dataset_prepared_path: last_run_prepared
|
278 |
+
val_set_size: 0.002
|
279 |
+
|
280 |
+
output_dir: ./Einstein-v6.1-Llama3-8B-model
|
281 |
+
|
282 |
+
sequence_len: 8192
|
283 |
+
sample_packing: true
|
284 |
+
pad_to_sequence_len: true
|
285 |
+
eval_sample_packing: false
|
286 |
+
|
287 |
+
wandb_project: Einstein
|
288 |
+
wandb_entity:
|
289 |
+
wandb_watch:
|
290 |
+
wandb_name: Einstein-v6.1-Llama3-2-epoch
|
291 |
+
wandb_log_model:
|
292 |
+
hub_model_id: Weyaxi/Einstein-v6.1-Llama3-8B
|
293 |
+
|
294 |
+
save_safetensors: true
|
295 |
+
|
296 |
+
gradient_accumulation_steps: 4
|
297 |
+
micro_batch_size: 1
|
298 |
+
num_epochs: 2
|
299 |
+
optimizer: adamw_bnb_8bit # look
|
300 |
+
lr_scheduler: cosine
|
301 |
+
learning_rate: 0.000005 # look
|
302 |
+
|
303 |
+
train_on_inputs: false
|
304 |
+
group_by_length: false
|
305 |
+
bf16: true
|
306 |
+
fp16: false
|
307 |
+
tf32: false
|
308 |
+
|
309 |
+
gradient_checkpointing: true
|
310 |
+
early_stopping_patience:
|
311 |
+
resume_from_checkpoint:
|
312 |
+
local_rank:
|
313 |
+
logging_steps: 1
|
314 |
+
xformers_attention:
|
315 |
+
flash_attention: true
|
316 |
+
|
317 |
+
warmup_steps: 10
|
318 |
+
evals_per_epoch: 2
|
319 |
+
eval_table_size:
|
320 |
+
eval_table_max_new_tokens: 128
|
321 |
+
saves_per_epoch: 2
|
322 |
+
debug:
|
323 |
+
|
324 |
+
deepspeed: zero3_bf16_cpuoffload_params.json
|
325 |
+
weight_decay: 0.0
|
326 |
+
fsdp:
|
327 |
+
fsdp_config:
|
328 |
+
special_tokens:
|
329 |
+
bos_token: "<s>"
|
330 |
+
eos_token: "<|im_end|>"
|
331 |
+
unk_token: "<unk>"
|
332 |
+
pad_token: <|end_of_text|> # changed
|
333 |
+
tokens:
|
334 |
+
- "<|im_start|>"
|
335 |
+
```
|
336 |
+
</details><br>
|
337 |
+
|
338 |
+
# π¬ Prompt Template
|
339 |
+
|
340 |
+
You can use ChatML prompt template while using the model:
|
341 |
+
|
342 |
+
### ChatML
|
343 |
+
|
344 |
+
```
|
345 |
+
<|im_start|>system
|
346 |
+
{system}<|im_end|>
|
347 |
+
<|im_start|>user
|
348 |
+
{user}<|im_end|>
|
349 |
+
<|im_start|>assistant
|
350 |
+
{asistant}<|im_end|>
|
351 |
+
```
|
352 |
+
|
353 |
+
This prompt template is available as a [chat template](https://huggingface.co/docs/transformers/main/chat_templating), which means you can format messages using the
|
354 |
+
`tokenizer.apply_chat_template()` method:
|
355 |
+
|
356 |
+
```python
|
357 |
+
messages = [
|
358 |
+
{"role": "system", "content": "You are helpful AI asistant."},
|
359 |
+
{"role": "user", "content": "Hello!"}
|
360 |
+
]
|
361 |
+
gen_input = tokenizer.apply_chat_template(message, return_tensors="pt")
|
362 |
+
model.generate(**gen_input)
|
363 |
+
```
|
364 |
+
|
365 |
+
# π Datasets used in this model
|
366 |
+
|
367 |
+
The datasets used to train this model are listed in the metadata section of the model card.
|
368 |
+
|
369 |
+
Please note that certain datasets mentioned in the metadata may have undergone filtering based on various criteria.
|
370 |
+
|
371 |
+
The results of this filtering process and its outcomes are in the data folder of this repository:
|
372 |
+
|
373 |
+
[Weyaxi/Einstein-v6.1-Llama3-8B/data](https://huggingface.co/Weyaxi/Einstein-v6.1-Llama3-8B/tree/main/data)
|
374 |
+
|
375 |
+
# π Quantizationed versions
|
376 |
+
|
377 |
+
## GGUF [@bartowski](https://huggingface.co/bartowski)
|
378 |
+
|
379 |
+
- https://huggingface.co/bartowski/Einstein-v6.1-Llama3-8B-GGUF
|
380 |
+
|
381 |
+
## ExLlamaV2 [@bartowski](https://huggingface.co/bartowski)
|
382 |
+
|
383 |
+
- https://huggingface.co/bartowski/Einstein-v6.1-Llama3-8B-exl2
|
384 |
+
|
385 |
+
## AWQ [@solidrust](https://huggingface.co/solidrust)
|
386 |
+
|
387 |
+
- https://huggingface.co/solidrust/Einstein-v6.1-Llama3-8B-AWQ
|
388 |
+
|
389 |
+
# π― [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
390 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_Weyaxi__Einstein-v6.1-Llama3-8B)
|
391 |
+
|
392 |
+
| Metric |Value|
|
393 |
+
|---------------------------------|----:|
|
394 |
+
|Avg. |68.60|
|
395 |
+
|AI2 Reasoning Challenge (25-Shot)|62.46|
|
396 |
+
|HellaSwag (10-Shot) |82.41|
|
397 |
+
|MMLU (5-Shot) |66.19|
|
398 |
+
|TruthfulQA (0-shot) |55.10|
|
399 |
+
|Winogrande (5-shot) |79.32|
|
400 |
+
|GSM8k (5-shot) |66.11|
|
401 |
+
|
402 |
+
# π€ Additional information about training
|
403 |
+
|
404 |
+
This model is full fine-tuned for 2 epoch.
|
405 |
+
|
406 |
+
Total number of steps was 2026.
|
407 |
+
|
408 |
+
<details><summary>Loss graph</summary>
|
409 |
+
|
410 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/6468ce47e134d050a58aa89c/Ycs7ZpoqmxFt0u9rybCO1.png)
|
411 |
+
|
412 |
+
</details><br>
|
413 |
+
|
414 |
+
# π€ Acknowledgments
|
415 |
+
|
416 |
+
Thanks to [sablo.ai](https://sablo.ai) for sponsoring this model.
|
417 |
+
|
418 |
+
Thanks to all the dataset authors mentioned in the datasets section.
|
419 |
+
|
420 |
+
Thanks to [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) for making the repository I used to make this model.
|
421 |
+
|
422 |
+
Thanks to all open source AI community.
|
423 |
+
|
424 |
+
[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
|
425 |
+
|
426 |
+
If you would like to support me:
|
427 |
+
|
428 |
+
[β Buy Me a Coffee](https://www.buymeacoffee.com/weyaxi)
|