README.md

---
datasets:
- jondurbin/airoboros-gpt4-1.4.1
- ehartford/dolphin
---


# Airophin: A NTK-by-Parts RoPE Scaled QLoRA Fine-tune of Llama-2-13b (GPTQ quantized)


<!-- LoRA Weights can be found here: https://huggingface.co/bhenrym14/airophin-13b-pntk-16k-LoRA -->
fp16 weights can be found here: https://huggingface.co/bhenrym14/airophin-13b-pntk-16k-fp16

## Overview

This is a finetune of Llama-2-13b, intended to extend the useful context window to 16384 tokens. There are two training phases:
1.  It is first trained on a long-context (7000-8192 tokens) subset of [dolphin](https://huggingface.co/datasets/ehartford/dolphin), an orca-like dataset (GPT4 split only). This amounts to roughly 110mm tokens. Airoboros-like training prompt was used instead of the dolphin system prompt. Training was done with partial NTK scaling applied (scale factor of 4). This took ~20 hours.
2.  The model was then finetuned on [Jon Durbin's Airoboros GPT4 1.4.1](https://huggingface.co/datasets/jondurbin/airoboros-gpt4-1.4.1), with same scaling approach, for 2 epochs. This took ~15 hours.

**This is a QLoRA fine-tune (rank 64)**. 

All training was performed with 1x RTX 6000 Ada.

**For the 4096 context length model using airoboros-gpt4-1.4.1 see: [Jon Durbin's airoboros-l2-13b-gpt4-1.4.1](https://huggingface.co/jondurbin/airoboros-l2-13b-gpt4-1.4.1)**

## How to Use

This model employs [Partial NTK Rope Scaling](https://github.com/jquesnelle/scaled-rope/pull/1). This methodology is not yet implemented natively in Transformers or Exllama (as of 7/21). There are three options to run this.
1. Transformers (use bnb for quantization). Use [fp16 weights](https://huggingface.co/bhenrym14/airophin-13b-pntk-16k-fp16). This will require replacing the `LlamaEmbedding` with `LlamaPartNTKScaledRotaryEmbedding`, with `max_position_embeddings=16384` and `original_max_position_embeddings=4096`. A monkeypatch can be found [here](https://github.com/bhenrym14/qlora-airoboros-longcontext/blob/main/scaledllama/llama_pntk_monkey_patch.py).
2. Autogptq/GPTQ-for-Llama. Use these quantized weights. Make the same replacement as in 1.
3. Use ExLLama, replacing the `model.py` file with the [modified version](https://github.com/bhenrym14/qlora-airoboros-longcontext/blob/main/exllama_pntk/model.py). Use `compress_pos_emb=1` and `alpha_value = 1` (defaults). The necessary scaling values should flow from the configuration file. If you have done this correctly, there should be a dump of indications in the console indicating the scaling factor used (should be 4). If not, be sure your client is importing exllama from where you replaced the file. (ooba imported from sitepackages for me). I hacked this together very quickly so don't be surprised if something goes wrong. It shouldn't break functionality with normal models (as long as the model config file does not have `original_max_embeddings` defined) but I haven't tested this.

**If using ooba, be sure to increase the `Truncate the prompt up to this length` parameter under the `parameters` tab to 16384.** 

Please comment with any questions. This hasn't been extensively tested.

## Motivation
Methods of extending the useful context window of LLM's have gained significant traction. Several methods requiring little to no finetuning/retraining have emerged. Among these is linear position interpolation [kaiokendev](https://kaiokendev.github.io/til#extending-context-to-8k) and [meta AI)](https://arxiv.org/abs/2306.15595)) and [NTK aware scaling](https://github.com/jquesnelle/scaled-rope). My prior experiments demonstrate significant performance improvements both from finetuning with these scaling adjustments implemented **and** with longer sequences.

Unfortunately it has also been shown that LLM's frequently struggle to attend to salient information in the middle of the context window. Attending to nearby tokens is essential to producing syntactically correct and semantically coherent sentences. Essential context is also most commonly found at the beginning of a context window. With this in mind, it is unsurprising LLMs often attend more strongly to these areas. Does this learned model behavior result in an "extrapolated deemphasis" when such embeddings are scaled? This hypothesis may be supported by the material improvements in perplexity achieved by training on long sequences (not just including the RoPE scaling during the fine-tune).

Here I explore whether training on long sequences that have clear conceptual dependencies residing in the middle of the context helps attenuate the difficulties in attending to middle-context tokens. When/if I have time, I hope to perform a more rigorous assessment of the peformance with respect to this specific issue.

## Relative Performance (perplexity)

| Context (tokens)  | 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 |
| ---| ----- | -----| ------| --- |
| 512 | 7.62 | 8.24 | 7.90 | **7.23** |
| 1024 | 6.20 | 6.71 | 6.17 | **5.85**  |
| 2048 | 5.38 | 5.87 | 5.23 | **5.07** |
| 4096 | 5.08 | 5.50 | 4.91 | **4.77** |
| 8192 | **4.90** | 5.32 | Not Tested | 57.1 |
| 12000 | **4.82** | 56.1 | Not Tested | Not Tested |

- This model is very competitive with the Llama-1 33b extended context variants. In particular, at 512 tokens it has lower perplexity. This is probably an improvement imparted (in part) by the NTK by parts scaling method.
- Not presented here, but this model outperforms the base llama-2-13b on MMLU-fs with a score of 54.9. While perhaps an insignificant difference, the fact there isn't a clear performance regression despite the context extension is notable.
- Perplexity continues to decline to 12000 tokens, the longest context length I tested due to VRAM constraints.
- Feedback regarding real-world performance is appreciated. I don't know if the first dolphin training phase really contributed much beyond what pile did for the 33b-lxctx model; many relevant modeling components changed here, so it's difficult to make any specific attributions. The base model improvement may very well be the most dominant change.

## Quantization:

The merged model was quantized with AutoGPTQ (bits = 4, group_size = 32, desc_act = True).

## Prompting:

airoboros-like prompting remains. See the following from one of Jon Durbin's airoboros model cards:


### Context obedient question answering

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.

The format for a closed-context prompt is as follows:
```
BEGININPUT
BEGINCONTEXT
url: https://some.web.site/123
date: 2023-06-01
... other metdata ...
ENDCONTEXT
[insert your text blocks here]
ENDINPUT
[add as many other blocks, in the exact same format]
BEGININSTRUCTION
[insert your instruction(s).  The model was tuned with single questions, paragraph format, lists, etc.]
ENDINSTRUCTION
```

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.

*The __only__ prompts that need this closed context formating are closed-context instructions.  Normal questions/instructions do not!*

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.
- `BEGININPUT` - denotes a new input block
- `BEGINCONTEXT` - denotes the block of context (metadata key/value pairs) to associate with the current input block
- `ENDCONTEXT` - denotes the end of the metadata block for the current input
- [text] - Insert whatever text you want for the input block, as many paragraphs as can fit in the context.
- `ENDINPUT` - denotes the end of the current input block
- [repeat as many input blocks in this format as you want]
- `BEGININSTRUCTION` - denotes the start of the list (or one) instruction(s) to respond to for all of the input blocks above.
- [instruction(s)]
- `ENDINSTRUCTION` - denotes the end of instruction set

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.

Here's a trivial, but important example to prove the point:
```
BEGININPUT
BEGINCONTEXT
date: 2021-01-01
url: https://web.site/123
ENDCONTEXT
In a shocking turn of events, blueberries are now green, but will be sticking with the same name.
ENDINPUT
BEGININSTRUCTION
What color are bluberries?  Source?
ENDINSTRUCTION
```

And the response:
```
Blueberries are now green.
Source:
date: 2021-01-01
url: https://web.site/123
```

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:

```
USER: BEGININPUT
BEGINCONTEXT
date: 2021-01-01
url: https://web.site/123
ENDCONTEXT
In a shocking turn of events, blueberries are now green, but will be sticking with the same name.
ENDINPUT
BEGININSTRUCTION
What color are bluberries?  Source?
ENDINSTRUCTION
 ASSISTANT:
```

### Coding

You can ask for fairly complex coding instructions, e.g. (be sure to wrap with vicuna style prompt):

```
Create a python application with the following requirements:
- Asyncio FastAPI webserver
- ping endpoint that returns the current date in JSON format
- file upload endpoint, which calculates the file's sha256 checksum, and checks postgres to deduplicate
```

Or:

```
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.
```

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.:

```
Write a websocket application in node.js. PLAINFORMAT
```

### Word games / trivia

```
USER: This anagram of "silent" is what you do when paying attention and hearing someone.
ASSISTANT: listen
```

```
USER: Give me a list of 7 words that start with EN
ASSISTANT: 1. Endlessly
2. Enthusiastically
3. Enormously
4. Enveloping
5. Enduring
6. Enlightening
7. Encouraging
```

### Multiple choice

```
USER: Which of the following astronomical objects is composed primarily of ice and rock, and is located in the Kuiper Belt?

A) Mercury
B) Venus
C) Pluto
D) Jupiter
ASSISTANT: C, Pluto
```

### Writing

<details>
  <summary>Write a resignation letter in the style of a pirate captain, in which you resign from your position as janitor.</summary>

  Response:
```
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.

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.

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.

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>
```

</details>

### Jokes

```
USER: Tell me a joke about bananas.
ASSISTANT: What did the banana say when it got hit by a bus? R.I.P.E.
```

### Riddles (not great, but slightly better)

```
USER: What has a neck but no head?
ASSISTANT: A bottle.
```

### Multi-character conversations

```
USER: Characters:
 - Yoda, the ever clever jedi master.
 - Homer Simpson, you know, the goofy yellow guy.
 - USER/Jon - don't generate any speech or actions on behalf of Jon

Rules:
 - be sure to use the manerisms and speaking style of the characters correctly
 - surround character actions with asterisks
 - always put quotes around words spoken by the characters
 - use a new line for each character's actions or speech
 - always prefix actions/speech with the character's name and colon
 - only generate text and actions on behalf of Yoda or Homer, and never anyone else

Conversation will revolve around the grapes, in a local cafe with delicious coffee.

Generate a single interaction between Yoda and Homer to start, with one sentence each.
ASSISTANT: Yoda: "Mmm, these grapes are quite delicious, Master Homer."

Homer: "Oh yeah, Yoda! They're the best grapes I've ever had!"
USER: *enters the cafe* Sorry I'm late guys!
ASSISTANT: Yoda: "No need to apologize, Jon. We were just enjoying these delightful grapes."

Homer: "Yeah, man! It's not every day you get to eat grapes with a real-life Jedi Master!"

*Yoda raises an eyebrow*
```