l3utterfly
commited on
Merge branch 'main' of https://huggingface.co/l3utterfly/phi-2-layla-v1
Browse files
README.md
CHANGED
@@ -1,149 +1,38 @@
|
|
1 |
---
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
- name: out
|
6 |
-
results: []
|
7 |
---
|
8 |
|
9 |
-
|
10 |
-
should probably proofread and complete it, then remove this comment. -->
|
11 |
|
12 |
-
|
13 |
-
<details><summary>See axolotl config</summary>
|
14 |
-
|
15 |
-
axolotl version: `0.4.0`
|
16 |
-
```yaml
|
17 |
-
base_model: /home/layla/src/text-generation-webui/models/phi-2
|
18 |
-
model_type: AutoModelForCausalLM
|
19 |
-
tokenizer_type: AutoTokenizer
|
20 |
-
|
21 |
-
load_in_8bit: false
|
22 |
-
load_in_4bit: false
|
23 |
-
strict: false
|
24 |
|
25 |
-
|
26 |
-
- path: /home/layla/src/Layla-datasets/datasets_formatted/base/dailydialog.topicalchat.teatime.openhermes.jsonl
|
27 |
-
ds_type: json # see other options below
|
28 |
-
type: sharegpt
|
29 |
-
conversation: vicuna_v1.1
|
30 |
|
31 |
-
|
32 |
-
|
33 |
-
|
|
|
|
|
34 |
|
35 |
-
|
36 |
-
|
37 |
-
|
|
|
|
|
|
|
38 |
|
39 |
-
|
40 |
-
sample_packing: true
|
41 |
-
pad_to_sequence_len: true
|
42 |
|
43 |
-
|
44 |
-
wandb_entity:
|
45 |
-
wandb_watch:
|
46 |
-
wandb_run_id:
|
47 |
-
wandb_log_model:
|
48 |
|
49 |
-
|
50 |
-
micro_batch_size: 2
|
51 |
-
num_epochs: 1
|
52 |
-
optimizer: adamw_bnb_8bit
|
53 |
-
lr_scheduler: cosine
|
54 |
-
learning_rate: 0.0000005
|
55 |
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
tf32: true
|
61 |
-
|
62 |
-
gradient_checkpointing: true
|
63 |
-
gradient_checkpointing_kwargs:
|
64 |
-
use_reentrant: True
|
65 |
-
early_stopping_patience:
|
66 |
-
resume_from_checkpoint:
|
67 |
-
local_rank:
|
68 |
-
logging_steps: 1
|
69 |
-
xformers_attention:
|
70 |
-
flash_attention: true
|
71 |
-
|
72 |
-
warmup_ratio: 0.05
|
73 |
-
eval_steps: 0.1
|
74 |
-
eval_sample_packing: true
|
75 |
-
save_steps: 300
|
76 |
-
debug:
|
77 |
-
deepspeed: /home/layla/src/Layla-datasets/axolotl/configs/deepspeed/zero2.json # multi-gpu only
|
78 |
-
weight_decay: 0.0
|
79 |
-
fsdp:
|
80 |
-
fsdp_config:
|
81 |
-
resize_token_embeddings_to_32x: true
|
82 |
-
special_tokens:
|
83 |
-
bos_token: "<|endoftext|>"
|
84 |
-
eos_token: "<|endoftext|>"
|
85 |
-
unk_token: "<|endoftext|>"
|
86 |
-
pad_token: "<|endoftext|>"
|
87 |
```
|
88 |
|
89 |
-
|
90 |
-
|
91 |
-
# out
|
92 |
-
|
93 |
-
This model was trained from scratch on the None dataset.
|
94 |
-
It achieves the following results on the evaluation set:
|
95 |
-
- Loss: 0.8072
|
96 |
-
|
97 |
-
## Model description
|
98 |
-
|
99 |
-
More information needed
|
100 |
-
|
101 |
-
## Intended uses & limitations
|
102 |
-
|
103 |
-
More information needed
|
104 |
-
|
105 |
-
## Training and evaluation data
|
106 |
-
|
107 |
-
More information needed
|
108 |
-
|
109 |
-
## Training procedure
|
110 |
-
|
111 |
-
### Training hyperparameters
|
112 |
-
|
113 |
-
The following hyperparameters were used during training:
|
114 |
-
- learning_rate: 5e-07
|
115 |
-
- train_batch_size: 2
|
116 |
-
- eval_batch_size: 2
|
117 |
-
- seed: 42
|
118 |
-
- distributed_type: multi-GPU
|
119 |
-
- num_devices: 5
|
120 |
-
- gradient_accumulation_steps: 4
|
121 |
-
- total_train_batch_size: 40
|
122 |
-
- total_eval_batch_size: 10
|
123 |
-
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
124 |
-
- lr_scheduler_type: cosine
|
125 |
-
- lr_scheduler_warmup_steps: 17
|
126 |
-
- num_epochs: 1
|
127 |
-
|
128 |
-
### Training results
|
129 |
-
|
130 |
-
| Training Loss | Epoch | Step | Validation Loss |
|
131 |
-
|:-------------:|:-----:|:----:|:---------------:|
|
132 |
-
| 0.9616 | 0.0 | 1 | 1.0031 |
|
133 |
-
| 0.9489 | 0.1 | 372 | 0.8825 |
|
134 |
-
| 0.987 | 0.2 | 744 | 0.8487 |
|
135 |
-
| 0.818 | 0.3 | 1116 | 0.8313 |
|
136 |
-
| 0.8389 | 0.4 | 1488 | 0.8212 |
|
137 |
-
| 0.9015 | 0.5 | 1860 | 0.8146 |
|
138 |
-
| 0.8237 | 0.6 | 2232 | 0.8108 |
|
139 |
-
| 0.7562 | 0.7 | 2604 | 0.8088 |
|
140 |
-
| 0.8776 | 0.8 | 2976 | 0.8078 |
|
141 |
-
| 0.8703 | 0.9 | 3348 | 0.8072 |
|
142 |
-
|
143 |
-
|
144 |
-
### Framework versions
|
145 |
-
|
146 |
-
- Transformers 4.39.0.dev0
|
147 |
-
- Pytorch 2.2.0
|
148 |
-
- Datasets 2.17.1
|
149 |
-
- Tokenizers 0.15.0
|
|
|
1 |
---
|
2 |
+
license: mit
|
3 |
+
language:
|
4 |
+
- en
|
|
|
|
|
5 |
---
|
6 |
|
7 |
+
# Model Card
|
|
|
8 |
|
9 |
+
### Model Description
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
+
Phi-2 fine-tuned by the OpenHermes 2.5 dataset optimised for multi-turn conversation and character impersonation.
|
|
|
|
|
|
|
|
|
12 |
|
13 |
+
The dataset has been pre-processed by doing the following:
|
14 |
+
1. remove all refusals
|
15 |
+
2. remove any mention of AI assistant
|
16 |
+
3. split any multi-turn dialog generated in the dataset into multi-turn conversations records
|
17 |
+
4. added nfsw generated conversations from the Teatime dataset
|
18 |
|
19 |
+
- **Developed by:** l3utterfly
|
20 |
+
- **Funded by:** Layla Network
|
21 |
+
- **Model type:** Phi-2
|
22 |
+
- **Language(s) (NLP):** English
|
23 |
+
- **License:** MIT
|
24 |
+
- **Finetuned from model:** Phi-2
|
25 |
|
26 |
+
## Uses
|
|
|
|
|
27 |
|
28 |
+
Base model used by Layla - the offline personal assistant: https://www.layla-network.ai
|
|
|
|
|
|
|
|
|
29 |
|
30 |
+
Help & support: https://discord.gg/x546YJ6nYC
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
+
Prompt:
|
33 |
+
```
|
34 |
+
USER:
|
35 |
+
ASSISTANT:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
```
|
37 |
|
38 |
+
[<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)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|