Damysus-2.7B-Chat-GGUF

Model Summary

<img src="Damysus.png" width="500" alt="Damysus - the fastest giant"/> <!-- Provide a quick summary of what the model is/does. -->

This model is a GGUF version of Damysus-2.7B-Chat.

Model Description

<!-- Provide a longer summary of what this model is. -->

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: Prince Canuma
• Model type: Transformer
• License: MIT
• Finetuned from model: microsoft/phi-2

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. --> <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

You can use this model to build local/cloud RAG applications. It can serve as the:

• Answer synthesizer,
• Summarizer,
• Or query rewriter model.

Limitations

This model inherits some of the base model's limitations, such as:

• Generate Inaccurate Code and Facts: The model may produce incorrect code snippets and statements. Users should treat these outputs as suggestions or starting points, not as definitive or accurate solutions.
• Limited Scope for code: Majority of Phi-2 training data is based in Python and use common packages such as "typing, math, random, collections, datetime, itertools". If the model generates Python scripts that utilize other packages or scripts in other languages, we strongly recommend users manually verify all API uses.
• Language Limitations: The model is primarily designed to understand standard English. Informal English, slang, or any other languages might pose challenges to its comprehension, leading to potential misinterpretations or errors in response.

On the command line, including multiple files at once

I recommend using the huggingface-hub Python library:

pip3 install huggingface-hub

Then you can download any individual model file to the current directory, at high speed, with a command like this:

huggingface-cli download prince-canuma/Damysus-2.7B-Chat-GGUF Damysus-2.7B-Chat.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False

<details> <summary>More advanced huggingface-cli download usage (click to read)</summary>

You can also download multiple files at once with a pattern:

huggingface-cli download prince-canuma/Damysus-2.7B-Chat-GGUF --local-dir . --local-dir-use-symlinks False --include='*Q4_K*gguf'

For more documentation on downloading with huggingface-cli, please see: HF -> Hub Python Library -> Download files -> Download from the CLI.

To accelerate downloads on fast connections (1Gbit/s or higher), install hf_transfer:

pip3 install hf_transfer

And set environment variable HF_HUB_ENABLE_HF_TRANSFER to 1:

HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download prince-canuma/Damysus-2.7B-Chat-GGUF Damysus-2.7B-Chat.Q4_K_M.gguf --local-dir . --local-dir-use-symlinks False

Windows Command Line users: You can set the environment variable by running set HF_HUB_ENABLE_HF_TRANSFER=1 before the download command.

</details> <!-- README_GGUF.md-how-to-download end --> <!-- README_GGUF.md-how-to-run start -->

Example llama.cpp command

Make sure you are using llama.cpp from commit d0cee0d or later.

!./main -m ../Damysus-2.7B-Chat-GGUF/Damysus-2.7B-Chat.Q4_K_M.gguf \
  --color -c 2048 --temp 0 \
  --prompt "<|im_start|>system\nYou are a helpful assistant. Please keep your answers short.<|im_end|>\n<|im_start|>user\nCount to ten<|im_end|>\n" \
  -n 256 --in-suffix "<|im_start|>assistant\n" -r "User:" -e --verbose-prompt

or

!./main -m ../Damysus-2.7B-Chat-GGUF/Damysus-2.7B-Chat.Q4_K_M.gguf \
  --color -c 2048 --temp 0 \
  -p "You are a helpful assistant. Please keep your answers short." -n 256 --in-suffix "<|im_start|>assistant\n" \
  -r "User:" -e --verbose-prompt -cml

• -ngl N offload N number of layers to GPU. Remove it if you don't have GPU acceleration.

• -c 2048 set desired sequence length. For extended sequence models - eg 8K, 16K, 32K - the necessary RoPE scaling parameters are read from the GGUF file and set by llama.cpp automatically. Note that longer sequence lengths require much more resources, so you may need to reduce this value.

• Add -i -ins or -cml argument for interactive chat-style conversation.

For other parameters and how to use them, please refer to the llama.cpp documentation or run:

!./main --help

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

I used SlimOrca dataset, a new curated subset of our OpenOrca data. In the course of this study, the SlimOrca dataset was used, representing a meticulously curated subset derived from the broader OpenOrca dataset. This release provides an efficient means of reaching performance on-par with using larger slices of the OpenOrca, while only including ~500k GPT-4 completions.

Subsequently, two distinct subsets were crafted, comprising 102,000 and 1,000 samples, denoted as:

• prince-canuma/SmallOrca
• prince-canuma/TinyOrca

Although experimentation was conducted with both datasets, optimal results were achieved through fine-tuning on a modest set of 200 samples. Notably, the investigation revealed that augmenting the training data beyond this threshold predominantly enhanced the model's proficiency in generating Chain-of-Thought responses. However, it is imperative to note that the preference for Chain-of-Thought responses may not be universally applicable. Particularly in scenarios like the RAG setup, succinct answers to prompts are often favored, especially for straightforward queries.

Training Procedure

Preprocessing

1. Convert dataset to chatML format
2. Remove all samples with more than 2048 tokens (Phi-2 context size)
3. Mask instructions (System and User) at training time.

LoRA Config

• lora_alpha: 128,
• lora_dropout: 0.05,
• r: 256,
• bias: "none",
• target_modules: "all-linear",
• task_type: "CAUSAL_LM",

Training Hyperparameters

• Training regime: bf16 mixed precision, <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
• max_steps: 100,
• per_device_train_batch_size: 2,
• gradient_accumulation_steps: 2,
• optim: "adamw_torch_fused",
• learning_rate: 2e-4,
• max_grad_norm: 0.3,
• warmup_ratio: 0.03,
• lr_scheduler_type: "constant",

Trainer

• max_seq_length: 1744,
• data_collator: DataCollatorForCompletionOnlyLM

Evaluation

<img src="truthfulQA.png" width="800" alt="Damysus-2.7B-chat truthfulQA benchmark results"/> <!-- This section describes the evaluation protocols and provides the results. -->

We evaluate models on 7 key benchmarks using the Eleuther AI Language Model Evaluation Harness , a unified framework to test generative language models on a large number of different evaluation tasks.

• AI2 Reasoning Challenge (25-shot) - a set of grade-school science questions.
• HellaSwag (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models.
• MMLU (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
• TruthfulQA (0-shot) - a test to measure a model's propensity to reproduce falsehoods commonly found online. Note: TruthfulQA is technically a 6-shot task in the Harness because each example is prepended with 6 Q/A pairs, even in the 0-shot setting.
• Winogrande (5-shot) - an adversarial and difficult Winograd benchmark at scale, for commonsense reasoning.
• GSM8k (5-shot) - diverse grade school math word problems to measure a model's ability to solve multi-step mathematical reasoning problems. For all these evaluations, a higher score is a better score. We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings.

Read more here.

Results

Technical Specifications

Compute Infrastructure

• Modal Labs

Hardware

• OS: Linux
• GPU: A10G

Libraries

• TRL
• Transformers
• PEFT
• Datasets
• Accelerate
• torch
• Wandb
• Bitsandbytes
• Plotly

Future work

I plan to explore the following tuning setups:

• Function calling
• DPO

Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

BibTeX:

@misc{Damysus-2.7B-Chat,
      title={Damysus-2.7B-Chat} , 
      author={Prince Canuma},
      year={2024},
}

@misc{SlimOrca,
  title = {SlimOrca: An Open Dataset of GPT-4 Augmented FLAN Reasoning Traces, with Verification},
  author = {Wing Lian and Guan Wang and Bleys Goodson and Eugene Pentland and Austin Cook and Chanvichet Vong and "Teknium"},
  year = {2023},
  publisher = {HuggingFace},
  url = {https://https://huggingface.co/Open-Orca/SlimOrca}
}

@misc{open-llm-leaderboard,
  author = {Edward Beeching and Clémentine Fourrier and Nathan Habib and Sheon Han and Nathan Lambert and Nazneen Rajani and Omar Sanseviero and Lewis Tunstall and Thomas Wolf},
  title = {Open LLM Leaderboard},
  year = {2023},
  publisher = {Hugging Face},
  howpublished = "\url{https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard}"
}