Awan LLM
commited on
Commit
•
f555a12
1
Parent(s):
7d5e6fa
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,148 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
Based on Meta-Llama-3-8b-Instruct, and is governed by Meta Llama 3 License agreement:
|
5 |
+
https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
|
6 |
+
|
7 |
+
|
8 |
+
We don't know how good this model is exactly in benchmarks since we have not benched this yet, but we think real prompts and usage is more telling anyways.
|
9 |
+
|
10 |
+
|
11 |
+
From our testing this model is:
|
12 |
+
|
13 |
+
- Less Refusals
|
14 |
+
- More Uncensored
|
15 |
+
- Follows requests better
|
16 |
+
- Can reply in requested formats better without adding unnecesary information
|
17 |
+
|
18 |
+
We are happy for anyone to try it out and give some feedback.
|
19 |
+
You can also try this model on our API at https://www.awanllm.com/
|
20 |
+
|
21 |
+
|
22 |
+
Training:
|
23 |
+
- 2048 sequence length, while the base model is 8192 sequence length. From testing it still performs the same 8192 context just fine.
|
24 |
+
- Trained on a modified and improved version of Cognitive Computations Eric Hartford's Dolphin dataset. https://huggingface.co/datasets/cognitivecomputations/dolphin
|
25 |
+
- Training duration is around 2 days on 2x RTX3090 on our own machine, using 4-bit loading and Qlora 64-rank 128-alpha resulting in ~2% trainable weights.
|
26 |
+
|
27 |
+
|
28 |
+
The goal for this model is to have the model less-censored and great at general tasks like the previous dolphin based models by Eric Hartford.
|
29 |
+
We started training this BEFORE they launched their own full weight trained Llama-3-8B-Dolphin-2.9 with their own curated datasets and the newer "Dolphin 2.9" dataset, but we think this model is still a unique take on Llama 3 8B Instruct and the dolphin dataset.
|
30 |
+
https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b
|
31 |
+
|
32 |
+
|
33 |
+
The difference with their dolphin 2.9 model is that we train this using Meta's new Llama 3 instruct format and not the regular ChatML format that Dolphin models are usually trained on.
|
34 |
+
This is because we think that it performed better using the format it was originally trained on.
|
35 |
+
|
36 |
+
Instruct format:
|
37 |
+
```
|
38 |
+
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
|
39 |
+
|
40 |
+
{{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|>
|
41 |
+
|
42 |
+
{{ user_message_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
|
43 |
+
|
44 |
+
{{ model_answer_1 }}<|eot_id|><|start_header_id|>user<|end_header_id|>
|
45 |
+
|
46 |
+
{{ user_message_2 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
|
47 |
+
```
|
48 |
+
|
49 |
+
|
50 |
+
Quants:
|
51 |
+
|
52 |
+
AWQ: https://huggingface.co/AwanLLM/Meta-Llama-3-8B-Instruct-Dolfin-AWQ
|
53 |
+
|
54 |
+
GGUF: https://huggingface.co/AwanLLM/Meta-Llama-3-8B-Instruct-Dolfin-v0.1-GGUF
|
55 |
+
|
56 |
+
FP16: https://huggingface.co/AwanLLM/Meta-Llama-3-8B-Instruct-Dolfin
|
57 |
+
|
58 |
+
Exllamav2:
|
59 |
+
|
60 |
+
4bpw: https://huggingface.co/AwanLLM/Meta-Llama-3-8B-Instruct-Dolfin-v0.1-exl2-h8-4bpw-exl2
|
61 |
+
|
62 |
+
8bpw: https://huggingface.co/AwanLLM/Meta-Llama-3-8B-Instruct-Dolfin-v0.1-exl2-h8-8bpw-exl2
|
63 |
+
|
64 |
+
|
65 |
+
[<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)
|
66 |
+
|
67 |
+
Axolotl Config:
|
68 |
+
```
|
69 |
+
base_model: Meta-Llama-3-8B-Instruct
|
70 |
+
model_type: LlamaForCausalLM
|
71 |
+
tokenizer_type: AutoTokenizer
|
72 |
+
|
73 |
+
train_on_inputs: false
|
74 |
+
group_by_length: false
|
75 |
+
load_in_8bit: false
|
76 |
+
load_in_4bit: true
|
77 |
+
strict: false
|
78 |
+
sequence_len: 2048
|
79 |
+
bf16: true
|
80 |
+
fp16: false
|
81 |
+
tf32: false
|
82 |
+
flash_attention: true
|
83 |
+
|
84 |
+
# Data
|
85 |
+
datasets:
|
86 |
+
- path: flan1m-universal-uncensored-system-2048.jsonl
|
87 |
+
type:
|
88 |
+
system_prompt: ""
|
89 |
+
system_format: "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"
|
90 |
+
field_system: system
|
91 |
+
field_instruction: input
|
92 |
+
field_output: output
|
93 |
+
format: "{instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
94 |
+
no_input_format: "{instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
|
95 |
+
|
96 |
+
warmup_steps: 10
|
97 |
+
dataset_prepared_path: ./last_run_prepared
|
98 |
+
|
99 |
+
# Iterations
|
100 |
+
num_epochs: 1
|
101 |
+
saves_per_epoch: 4
|
102 |
+
|
103 |
+
# Evaluation
|
104 |
+
val_set_size: 0.01
|
105 |
+
eval_table_size:
|
106 |
+
eval_table_max_new_tokens:
|
107 |
+
eval_sample_packing: false
|
108 |
+
evals_per_epoch: 4
|
109 |
+
|
110 |
+
# LoRA
|
111 |
+
output_dir: ./qlora-out
|
112 |
+
adapter: qlora
|
113 |
+
lora_model_dir:
|
114 |
+
lora_r: 64
|
115 |
+
lora_alpha: 128
|
116 |
+
lora_dropout: 0.05
|
117 |
+
lora_target_linear: true
|
118 |
+
lora_fan_in_fan_out:
|
119 |
+
lora_target_modules:
|
120 |
+
save_safetensors: true
|
121 |
+
|
122 |
+
# Sampling
|
123 |
+
sample_packing: true
|
124 |
+
pad_to_sequence_len: true
|
125 |
+
|
126 |
+
# Batching
|
127 |
+
gradient_accumulation_steps: 32
|
128 |
+
micro_batch_size: 4
|
129 |
+
gradient_checkpointing: true
|
130 |
+
gradient_checkpointing_kwargs:
|
131 |
+
use_reentrant: true
|
132 |
+
|
133 |
+
# Optimizer
|
134 |
+
optimizer: paged_adamw_8bit
|
135 |
+
lr_scheduler: cosine
|
136 |
+
learning_rate: 0.0002
|
137 |
+
|
138 |
+
# Misc
|
139 |
+
early_stopping_patience:
|
140 |
+
resume_from_checkpoint:
|
141 |
+
logging_steps: 1
|
142 |
+
debug:
|
143 |
+
deepspeed: zero3_bf16.json
|
144 |
+
weight_decay: 0.1
|
145 |
+
special_tokens:
|
146 |
+
pad_token: <|end_of_text|>
|
147 |
+
```
|
148 |
+
|