Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,247 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
datasets:
|
3 |
+
- jondurbin/airoboros-gpt4-m2.0
|
4 |
+
- ehartford/dolphin
|
5 |
+
- shahules786/orca-chat
|
6 |
+
---
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
# Airophin: An Airoboros-Dolphin Extended Context QLoRA Fine-tune of Llama-2-13b (GPTQ weights)
|
11 |
+
|
12 |
+
|
13 |
+
fp16 Weights can be found here: https://huggingface.co/bhenrym14/airophin-v2-13b-PI-8k-fp16
|
14 |
+
<!-- GPTQ weights can be found here: https://huggingface.co/bhenrym14/airophin-13b-pntk-16k-GPTQ -->
|
15 |
+
|
16 |
+
## Overview
|
17 |
+
|
18 |
+
This is a finetune of Llama-2-13b, intended to extend the useful context window to 8192 tokens via position interpolation (PI). There are two training phases, but in this model I only perform the final finetune on the Airoboros m2.0 dataset.
|
19 |
+
1. I start with [OpenAssistant/llama2-13b-orca-8k-3319](https://huggingface.co/OpenAssistant/llama2-13b-orca-8k-3319). This model has been trained on a mix of orca-chat (dolphin derived), fanfics, and redpajama; the majority of the dataset is orca-chat, hence why I retain the airophin naming for this model.
|
20 |
+
2. The model was then finetuned on the merged Airoboros dataset (1.4.1 merged with 2.0) [Jon Durbin's Airoboros GPT4 m2.0](https://huggingface.co/datasets/jondurbin/airoboros-gpt4-m2.0), with same scaling approach, for 2 epochs.
|
21 |
+
|
22 |
+
**This is a (merged) QLoRA fine-tune (rank 64)**.
|
23 |
+
|
24 |
+
The finetune was performed with 1x RTX 6000 Ada.
|
25 |
+
|
26 |
+
|
27 |
+
## How to Use
|
28 |
+
|
29 |
+
This model employs linear RoPE scaling. Transformers now has native support (be sure to update). To use transformers/bnb, you'll want the fp16 weights from the repo linked above.
|
30 |
+
|
31 |
+
To use these GPTQ weights, use AutoGPTQ or exllama. For exllama be sure that `max_seq_len=8192` and `compress_pos_emb=2`. Please comment if you have issues with this.
|
32 |
+
|
33 |
+
Ooba use: Be sure to increase the `Truncate the prompt up to this length` parameter to 8192 to utilize the full context capabilities.
|
34 |
+
|
35 |
+
## Motivation
|
36 |
+
|
37 |
+
Previous experiments have demonstrated that orca-like datasets yield substantial performance improvements on numerous benchmarks. Additionally, the PI method of context extension requires finetuning to minimize performance impacts relative to the original (non context extended) model. My most successful models for context extension with PI methods employ a pretraining phase on long sequences, but due to the compute requirements, I have not scaled this to more than 200 iterations or so. Many groups (including OpenAssistant) have performed such training at scale. This model uses such a model as a starting point.
|
38 |
+
|
39 |
+
## Relative Performance (perplexity)
|
40 |
+
|
41 |
+
| Context (tokens) | bhenrym14/airophin-v2-13b-PI-8k-fp16 | bhenrym14/airophin-13b-pntk-16k-fp16| bhenrym14/airoboros-13b-gpt4-1.4.1-PI-8192-fp16 |bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16 | jondurbin/airoboros-l2-13b-gpt4-1.4.1 |
|
42 |
+
| --- | ---| ----- | -----| ------| --- |
|
43 |
+
| 512 | 7.38 | 7.62 | 8.24 | 7.90 | **7.23** |
|
44 |
+
| 1024 | 5.99 | 6.20 | 6.71 | 6.17 | **5.85** |
|
45 |
+
| 2048 | 5.22 | 5.38 | 5.87 | 5.23 | **5.07** |
|
46 |
+
| 4096 | 4.90 | 5.08 | 5.50 | 4.91 | **4.77** |
|
47 |
+
| 8192 | **4.71** | 4.90 | 5.32 | Not Tested | 57.1 |
|
48 |
+
| 12000 | 55 | **4.82** | 56.1 | Not Tested | Not Tested |
|
49 |
+
|
50 |
+
- This model is very competitive with the Llama-1 33b extended context variants. In fact, it outperforms bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16 everywhere <=8192 tokens. Do note however that 33b model is only trained on the 1.4.1 Airoboros dataset. Additionally this model only requires a PI factor of 2, whereas the 33b-16k llama1 model requires a factor of 8. It is clear from my experiments and those in the literature that higher factors pose larger challenges for performance recovery.
|
51 |
+
- Not presented here, but this model outperforms the base llama-2-13b on MMLU-fs with a score of 58.3. If this score ends up being be replicated on the HF LLM leaderboard, **this would place this model at 2nd or 3rd overall for MMLU among 13b models (and the #1 for extended context)**
|
52 |
+
- Feedback regarding real-world performance is appreciated. Llama2-13b is known to have repetition problems. Does the extensive training on top of the base model help ameliorate this tendency? Perplexity and MMLU are great, but the don't tell the whole story.
|
53 |
+
|
54 |
+
## Prompting:
|
55 |
+
|
56 |
+
This model was trained with airoboros-like prompting in the 2nd phase. See the following from one of Jon Durbin's airoboros model cards:
|
57 |
+
|
58 |
+
|
59 |
+
### Context obedient question answering
|
60 |
+
|
61 |
+
By obedient, I mean the model was trained to ignore what it thinks it knows, and uses the context to answer the question. The model was also tuned to limit the values to the provided context as much as possible to reduce hallucinations.
|
62 |
+
|
63 |
+
The format for a closed-context prompt is as follows:
|
64 |
+
```
|
65 |
+
BEGININPUT
|
66 |
+
BEGINCONTEXT
|
67 |
+
url: https://some.web.site/123
|
68 |
+
date: 2023-06-01
|
69 |
+
... other metdata ...
|
70 |
+
ENDCONTEXT
|
71 |
+
[insert your text blocks here]
|
72 |
+
ENDINPUT
|
73 |
+
[add as many other blocks, in the exact same format]
|
74 |
+
BEGININSTRUCTION
|
75 |
+
[insert your instruction(s). The model was tuned with single questions, paragraph format, lists, etc.]
|
76 |
+
ENDINSTRUCTION
|
77 |
+
```
|
78 |
+
|
79 |
+
It's also helpful to add "Don't make up answers if you don't know." to your instruction block to make sure if the context is completely unrelated it doesn't make something up.
|
80 |
+
|
81 |
+
*The __only__ prompts that need this closed context formating are closed-context instructions. Normal questions/instructions do not!*
|
82 |
+
|
83 |
+
I know it's a bit verbose and annoying, but after much trial and error, using these explicit delimiters helps the model understand where to find the responses and how to associate specific sources with it.
|
84 |
+
- `BEGININPUT` - denotes a new input block
|
85 |
+
- `BEGINCONTEXT` - denotes the block of context (metadata key/value pairs) to associate with the current input block
|
86 |
+
- `ENDCONTEXT` - denotes the end of the metadata block for the current input
|
87 |
+
- [text] - Insert whatever text you want for the input block, as many paragraphs as can fit in the context.
|
88 |
+
- `ENDINPUT` - denotes the end of the current input block
|
89 |
+
- [repeat as many input blocks in this format as you want]
|
90 |
+
- `BEGININSTRUCTION` - denotes the start of the list (or one) instruction(s) to respond to for all of the input blocks above.
|
91 |
+
- [instruction(s)]
|
92 |
+
- `ENDINSTRUCTION` - denotes the end of instruction set
|
93 |
+
|
94 |
+
It sometimes works without `ENDINSTRUCTION`, but by explicitly including that in the prompt, the model better understands that all of the instructions in the block should be responded to.
|
95 |
+
|
96 |
+
Here's a trivial, but important example to prove the point:
|
97 |
+
```
|
98 |
+
BEGININPUT
|
99 |
+
BEGINCONTEXT
|
100 |
+
date: 2021-01-01
|
101 |
+
url: https://web.site/123
|
102 |
+
ENDCONTEXT
|
103 |
+
In a shocking turn of events, blueberries are now green, but will be sticking with the same name.
|
104 |
+
ENDINPUT
|
105 |
+
BEGININSTRUCTION
|
106 |
+
What color are bluberries? Source?
|
107 |
+
ENDINSTRUCTION
|
108 |
+
```
|
109 |
+
|
110 |
+
And the response:
|
111 |
+
```
|
112 |
+
Blueberries are now green.
|
113 |
+
Source:
|
114 |
+
date: 2021-01-01
|
115 |
+
url: https://web.site/123
|
116 |
+
```
|
117 |
+
|
118 |
+
The prompt itself should be wrapped in the vicuna1.1 template if you aren't using fastchat with the conv-template vicuna_v1.1 as described:
|
119 |
+
|
120 |
+
```
|
121 |
+
USER: BEGININPUT
|
122 |
+
BEGINCONTEXT
|
123 |
+
date: 2021-01-01
|
124 |
+
url: https://web.site/123
|
125 |
+
ENDCONTEXT
|
126 |
+
In a shocking turn of events, blueberries are now green, but will be sticking with the same name.
|
127 |
+
ENDINPUT
|
128 |
+
BEGININSTRUCTION
|
129 |
+
What color are bluberries? Source?
|
130 |
+
ENDINSTRUCTION
|
131 |
+
ASSISTANT:
|
132 |
+
```
|
133 |
+
|
134 |
+
### Coding
|
135 |
+
|
136 |
+
You can ask for fairly complex coding instructions, e.g. (be sure to wrap with vicuna style prompt):
|
137 |
+
|
138 |
+
```
|
139 |
+
Create a python application with the following requirements:
|
140 |
+
- Asyncio FastAPI webserver
|
141 |
+
- ping endpoint that returns the current date in JSON format
|
142 |
+
- file upload endpoint, which calculates the file's sha256 checksum, and checks postgres to deduplicate
|
143 |
+
```
|
144 |
+
|
145 |
+
Or:
|
146 |
+
|
147 |
+
```
|
148 |
+
Write a multi-threaded TCP server in C that accepts a "GET [key]" input and "SET [key] [value]" input, and uses a binary tree to get and store the input values.
|
149 |
+
```
|
150 |
+
|
151 |
+
You can optionally add a single space and "PLAINFORMAT" at the end of your prompt to avoid backticks, explanations, etc. and just print the code, e.g.:
|
152 |
+
|
153 |
+
```
|
154 |
+
Write a websocket application in node.js. PLAINFORMAT
|
155 |
+
```
|
156 |
+
|
157 |
+
### Word games / trivia
|
158 |
+
|
159 |
+
```
|
160 |
+
USER: This anagram of "silent" is what you do when paying attention and hearing someone.
|
161 |
+
ASSISTANT: listen
|
162 |
+
```
|
163 |
+
|
164 |
+
```
|
165 |
+
USER: Give me a list of 7 words that start with EN
|
166 |
+
ASSISTANT: 1. Endlessly
|
167 |
+
2. Enthusiastically
|
168 |
+
3. Enormously
|
169 |
+
4. Enveloping
|
170 |
+
5. Enduring
|
171 |
+
6. Enlightening
|
172 |
+
7. Encouraging
|
173 |
+
```
|
174 |
+
|
175 |
+
### Multiple choice
|
176 |
+
|
177 |
+
```
|
178 |
+
USER: Which of the following astronomical objects is composed primarily of ice and rock, and is located in the Kuiper Belt?
|
179 |
+
|
180 |
+
A) Mercury
|
181 |
+
B) Venus
|
182 |
+
C) Pluto
|
183 |
+
D) Jupiter
|
184 |
+
ASSISTANT: C, Pluto
|
185 |
+
```
|
186 |
+
|
187 |
+
### Writing
|
188 |
+
|
189 |
+
<details>
|
190 |
+
<summary>Write a resignation letter in the style of a pirate captain, in which you resign from your position as janitor.</summary>
|
191 |
+
|
192 |
+
Response:
|
193 |
+
```
|
194 |
+
ASSISTANT: Captain Jack Sparrow here, as ye be knowin', has been forced to abandon me post as Janitor o' th' Festering Cesspool, which be a filthy hole o' water that we be callin' home.
|
195 |
+
|
196 |
+
Ye may be thinkin' what could ha' driven me to leave me duties an' take me leave o' this sorry establishment? Well, ye'd be right in thinkin' so, for it be th' greed o' them what runs th' place that be drivin' me away. They've been fattenin' their pockets on me back while I've been slavin' me guts out cleanin' up their messes.
|
197 |
+
|
198 |
+
No more, me hearties! Me time o' servin' th' rich an' powerful be done, an' I'm takin' me leave o' this stinkin' hole forever! Me destiny lies elsewhere, an' it be time I pursued it. Ye may think ye've seen th' last o' me, but ye'd be wrong, for th' winds o' change be blowin', an' me ship be ready to set sail on th' high seas o' justice.
|
199 |
+
|
200 |
+
So, farewell ye scurvy dogs, ye who've made me life a living hell. I leave ye with me favorite saying: "ARRRGGGHHH, ye scurvy dogs!" An' remember, ye ain't never gonna see me comin', for me shadow be castin' long an' dark on th' horizon</summary>
|
201 |
+
```
|
202 |
+
|
203 |
+
</details>
|
204 |
+
|
205 |
+
### Jokes
|
206 |
+
|
207 |
+
```
|
208 |
+
USER: Tell me a joke about bananas.
|
209 |
+
ASSISTANT: What did the banana say when it got hit by a bus? R.I.P.E.
|
210 |
+
```
|
211 |
+
|
212 |
+
### Riddles (not great, but slightly better)
|
213 |
+
|
214 |
+
```
|
215 |
+
USER: What has a neck but no head?
|
216 |
+
ASSISTANT: A bottle.
|
217 |
+
```
|
218 |
+
|
219 |
+
### Multi-character conversations
|
220 |
+
|
221 |
+
```
|
222 |
+
USER: Characters:
|
223 |
+
- Yoda, the ever clever jedi master.
|
224 |
+
- Homer Simpson, you know, the goofy yellow guy.
|
225 |
+
- USER/Jon - don't generate any speech or actions on behalf of Jon
|
226 |
+
|
227 |
+
Rules:
|
228 |
+
- be sure to use the manerisms and speaking style of the characters correctly
|
229 |
+
- surround character actions with asterisks
|
230 |
+
- always put quotes around words spoken by the characters
|
231 |
+
- use a new line for each character's actions or speech
|
232 |
+
- always prefix actions/speech with the character's name and colon
|
233 |
+
- only generate text and actions on behalf of Yoda or Homer, and never anyone else
|
234 |
+
|
235 |
+
Conversation will revolve around the grapes, in a local cafe with delicious coffee.
|
236 |
+
|
237 |
+
Generate a single interaction between Yoda and Homer to start, with one sentence each.
|
238 |
+
ASSISTANT: Yoda: "Mmm, these grapes are quite delicious, Master Homer."
|
239 |
+
|
240 |
+
Homer: "Oh yeah, Yoda! They're the best grapes I've ever had!"
|
241 |
+
USER: *enters the cafe* Sorry I'm late guys!
|
242 |
+
ASSISTANT: Yoda: "No need to apologize, Jon. We were just enjoying these delightful grapes."
|
243 |
+
|
244 |
+
Homer: "Yeah, man! It's not every day you get to eat grapes with a real-life Jedi Master!"
|
245 |
+
|
246 |
+
*Yoda raises an eyebrow*
|
247 |
+
```
|