model card #1
Browse files
README.md
ADDED
@@ -0,0 +1,269 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
tags:
|
4 |
+
- axolotl
|
5 |
+
- generated_from_trainer
|
6 |
+
- Mistral
|
7 |
+
- instruct
|
8 |
+
- finetune
|
9 |
+
- chatml
|
10 |
+
- gpt4
|
11 |
+
- synthetic data
|
12 |
+
- science
|
13 |
+
- physics
|
14 |
+
- chemistry
|
15 |
+
- biology
|
16 |
+
- math
|
17 |
+
base_model: mistralai/Mistral-7B-v0.1
|
18 |
+
datasets:
|
19 |
+
- allenai/ai2_arc
|
20 |
+
- camel-ai/physics
|
21 |
+
- camel-ai/chemistry
|
22 |
+
- camel-ai/biology
|
23 |
+
- camel-ai/math
|
24 |
+
- metaeval/reclor
|
25 |
+
- openbookqa
|
26 |
+
- mandyyyyii/scibench
|
27 |
+
- derek-thomas/ScienceQA
|
28 |
+
- TIGER-Lab/ScienceEval
|
29 |
+
- jondurbin/airoboros-3.2
|
30 |
+
- LDJnr/Capybara
|
31 |
+
- Cot-Alpaca-GPT4-From-OpenHermes-2.5
|
32 |
+
- STEM-AI-mtl/Electrical-engineering
|
33 |
+
- knowrohit07/saraswati-stem
|
34 |
+
- sablo/oasst2_curated
|
35 |
+
- lmsys/lmsys-chat-1m
|
36 |
+
- TIGER-Lab/MathInstruct
|
37 |
+
- bigbio/med_qa
|
38 |
+
- meta-math/MetaMathQA-40K
|
39 |
+
- openbookqa
|
40 |
+
- piqa
|
41 |
+
- metaeval/reclor
|
42 |
+
- derek-thomas/ScienceQA
|
43 |
+
- scibench
|
44 |
+
- sciq
|
45 |
+
- Open-Orca/SlimOrca
|
46 |
+
- migtissera/Synthia-v1.3
|
47 |
+
- TIGER-Lab/ScienceEval
|
48 |
+
- allenai/WildChat
|
49 |
+
- microsoft/orca-math-word-problems-200k
|
50 |
+
- openchat/openchat_sharegpt4_dataset
|
51 |
+
- teknium/GPTeacher-General-Instruct
|
52 |
+
- m-a-p/CodeFeedback-Filtered-Instruction
|
53 |
+
---
|
54 |
+
|
55 |
+
![image](image)
|
56 |
+
# 🔬 Einstein-v5-v0.2-7B
|
57 |
+
|
58 |
+
This model is a full fine-tuned version of [alpindale/Mistral-7B-v0.2-hf](https://huggingface.co/alpindale/Mistral-7B-v0.2-hf) on diverse datasets.
|
59 |
+
|
60 |
+
This model is finetuned using `8xRTX3090` + `1xRTXA6000` using [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl).
|
61 |
+
|
62 |
+
This model's training was sponsored by [sablo.ai](https://sablo.ai).
|
63 |
+
|
64 |
+
<details><summary>See axolotl config</summary>
|
65 |
+
|
66 |
+
axolotl version: `0.4.0`
|
67 |
+
```yaml
|
68 |
+
base_model: alpindale/Mistral-7B-v0.2-hf
|
69 |
+
model_type: MistralForCausalLM
|
70 |
+
tokenizer_type: LlamaTokenizer
|
71 |
+
is_mistral_derived_model: true
|
72 |
+
|
73 |
+
load_in_8bit: false
|
74 |
+
load_in_4bit: false
|
75 |
+
strict: false
|
76 |
+
|
77 |
+
chat_template: chatml
|
78 |
+
datasets:
|
79 |
+
- path: data/merged_all.json
|
80 |
+
ds_type: json
|
81 |
+
type: alpaca
|
82 |
+
conversation: chatml
|
83 |
+
|
84 |
+
- path: data/gpteacher-instruct-special-alpaca.json
|
85 |
+
ds_type: json
|
86 |
+
type: gpteacher
|
87 |
+
conversation: chatml
|
88 |
+
|
89 |
+
- path: data/capybara_sharegpt.json
|
90 |
+
ds_type: json
|
91 |
+
type: sharegpt
|
92 |
+
conversation: chatml
|
93 |
+
|
94 |
+
- path: data/synthia-v1.3_sharegpt_12500.json
|
95 |
+
ds_type: json
|
96 |
+
type: sharegpt
|
97 |
+
conversation: chatml
|
98 |
+
|
99 |
+
- path: data/cot_alpaca_gpt4_extracted_openhermes_2.5_sharegpt.json
|
100 |
+
ds_type: json
|
101 |
+
type: sharegpt
|
102 |
+
conversation: chatml
|
103 |
+
|
104 |
+
- path: data/slimorca_dedup_filtered_95k_sharegpt.json
|
105 |
+
ds_type: json
|
106 |
+
type: sharegpt
|
107 |
+
conversation: chatml
|
108 |
+
|
109 |
+
- path: data/airoboros_3.2_without_contextual_slimorca_orca_sharegpt.json
|
110 |
+
ds_type: json
|
111 |
+
type: sharegpt
|
112 |
+
conversation: chatml
|
113 |
+
|
114 |
+
- path: data/allenai_wild_chat_gpt4_english_toxic_random_half_4k_sharegpt.json
|
115 |
+
ds_type: json
|
116 |
+
type: sharegpt
|
117 |
+
strict: false
|
118 |
+
conversation: chatml
|
119 |
+
|
120 |
+
- path: data/pippa_bagel_repo_3k_sharegpt.json
|
121 |
+
ds_type: json
|
122 |
+
type: sharegpt
|
123 |
+
conversation: chatml
|
124 |
+
|
125 |
+
- path: data/gpt4_data_lmys_1m_sharegpt.json
|
126 |
+
ds_type: json
|
127 |
+
type: sharegpt
|
128 |
+
conversation: chatml
|
129 |
+
|
130 |
+
- path: data/sharegpt_gpt4_english.json
|
131 |
+
ds_type: json
|
132 |
+
type: sharegpt
|
133 |
+
conversation: chatml
|
134 |
+
|
135 |
+
dataset_prepared_path: last_run_prepared
|
136 |
+
# val_set_size: 0.005
|
137 |
+
val_set_size: 0.0
|
138 |
+
|
139 |
+
do_bench_eval: true
|
140 |
+
|
141 |
+
output_dir: ./Einstein-v5-Mistral-v0.2-beta-model
|
142 |
+
|
143 |
+
sequence_len: 8192
|
144 |
+
sample_packing: true
|
145 |
+
pad_to_sequence_len: true
|
146 |
+
eval_sample_packing: false
|
147 |
+
|
148 |
+
wandb_project: Einstein
|
149 |
+
wandb_entity:
|
150 |
+
wandb_watch:
|
151 |
+
wandb_name:
|
152 |
+
wandb_log_model:
|
153 |
+
hub_model_id: Weyaxi/Einstein-v5-Mistral-v0.2-beta
|
154 |
+
|
155 |
+
save_safetensors: true
|
156 |
+
|
157 |
+
gradient_accumulation_steps: 4
|
158 |
+
micro_batch_size: 1
|
159 |
+
num_epochs: 2
|
160 |
+
optimizer: adamw_bnb_8bit
|
161 |
+
lr_scheduler: cosine
|
162 |
+
learning_rate: 0.000005
|
163 |
+
|
164 |
+
train_on_inputs: false
|
165 |
+
group_by_length: false
|
166 |
+
bf16: true
|
167 |
+
fp16: false
|
168 |
+
tf32: false
|
169 |
+
|
170 |
+
gradient_checkpointing: true
|
171 |
+
early_stopping_patience:
|
172 |
+
resume_from_checkpoint:
|
173 |
+
local_rank:
|
174 |
+
logging_steps: 1
|
175 |
+
xformers_attention:
|
176 |
+
flash_attention: true
|
177 |
+
|
178 |
+
warmup_steps: 10
|
179 |
+
evals_per_epoch: 3 # changed
|
180 |
+
eval_table_size:
|
181 |
+
eval_table_max_new_tokens: 128
|
182 |
+
saves_per_epoch: 3 # changed
|
183 |
+
debug:
|
184 |
+
|
185 |
+
deepspeed: zero3_bf16.json
|
186 |
+
weight_decay: 0.0
|
187 |
+
fsdp:
|
188 |
+
fsdp_config:
|
189 |
+
special_tokens:
|
190 |
+
bos_token: "<s>"
|
191 |
+
eos_token: "<|im_end|>"
|
192 |
+
unk_token: "<unk>"
|
193 |
+
tokens:
|
194 |
+
- "<|im_start|>"
|
195 |
+
```
|
196 |
+
|
197 |
+
</details><br>
|
198 |
+
|
199 |
+
# 💬 Prompt Template
|
200 |
+
|
201 |
+
You can use this prompt template while using the model:
|
202 |
+
|
203 |
+
### ChatML
|
204 |
+
|
205 |
+
```
|
206 |
+
<|im_start|>system
|
207 |
+
{system}<|im_end|>
|
208 |
+
<|im_start|>user
|
209 |
+
{user}<|im_end|>
|
210 |
+
<|im_start|>assistant
|
211 |
+
{asistant}<|im_end|>
|
212 |
+
```
|
213 |
+
|
214 |
+
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
|
215 |
+
`tokenizer.apply_chat_template()` method:
|
216 |
+
|
217 |
+
```python
|
218 |
+
messages = [
|
219 |
+
{"role": "system", "content": "You are helpful AI asistant."},
|
220 |
+
{"role": "user", "content": "Hello!"}
|
221 |
+
]
|
222 |
+
gen_input = tokenizer.apply_chat_template(message, return_tensors="pt")
|
223 |
+
model.generate(**gen_input)
|
224 |
+
```
|
225 |
+
|
226 |
+
# 🔄 Quantizationed versions
|
227 |
+
|
228 |
+
Quantizationed versions of this model is currently not available. It will be out shortly :)
|
229 |
+
|
230 |
+
|
231 |
+
# 🎯 [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
232 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_Weyaxi__Einstein-v5-v0.2-7B)
|
233 |
+
|
234 |
+
| Metric |Value|
|
235 |
+
|---------------------------------|----:|
|
236 |
+
|Avg. |x|
|
237 |
+
|AI2 Reasoning Challenge (25-Shot)|x|
|
238 |
+
|HellaSwag (10-Shot) |x|
|
239 |
+
|MMLU (5-Shot) |x|
|
240 |
+
|TruthfulQA (0-shot) |x|
|
241 |
+
|Winogrande (5-shot) |x|
|
242 |
+
|GSM8k (5-shot) |x|
|
243 |
+
|
244 |
+
# 🤖 Additional information about training
|
245 |
+
|
246 |
+
This model is full fine-tuned for 1 epoch.
|
247 |
+
|
248 |
+
Total number of steps was 1124.
|
249 |
+
|
250 |
+
<details><summary>Loss graph</summary>
|
251 |
+
|
252 |
+
![image](image)
|
253 |
+
</details><br>
|
254 |
+
|
255 |
+
# 🤝 Acknowledgments
|
256 |
+
|
257 |
+
Thanks to [sablo.ai](https://sablo.ai) for sponsoring this model.
|
258 |
+
|
259 |
+
Thanks to all the dataset authors mentioned in the datasets section.
|
260 |
+
|
261 |
+
Thanks to [axolotl](https://github.com/OpenAccess-AI-Collective/axolotl) for making the repository I used to make this model.
|
262 |
+
|
263 |
+
Thanks to all open source AI community.
|
264 |
+
|
265 |
+
[<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)
|
266 |
+
|
267 |
+
If you would like to support me:
|
268 |
+
|
269 |
+
[☕ Buy Me a Coffee](https://www.buymeacoffee.com/weyaxi)
|