README.md

metadata

license: apache-2.0
language:
  - en
  - de
  - es
  - fr
tags:
  - sft
inference: false
datasets:
  - OpenAssistant/oasst1
  - databricks/databricks-dolly-15k

Chat & support: my new Discord server

Want to contribute? TheBloke's Patreon page

Open Assistant's Falcon 40B SFT MIX GGML

These files are GGCC format model files for Open Assistant's Falcon 40B SFT MIX.

These files will not work in llama.cpp, text-generation-webui or KoboldCpp.

GGCC is a new format created in a new fork of llama.cpp that introduced this new Falcon GGML-based support: cmp-nc/ggllm.cpp.

Currently these files will also not work with code that previously supported Falcon, such as LoLLMs Web UI and ctransformers. But support should be added soon.

Repositories available

• 2, 3, 4, 5, 6, 8-bit GGCT models for CPU+GPU inference
• Unquantised fp16 model in pytorch format, for GPU inference and for further conversions

Compatibility

To build cmp-nct's fork of llama.cpp with Falcon support plus CUDA acceleration, please try the following steps:

git clone https://github.com/cmp-nct/ggllm.cpp
cd ggllm.cpp
rm -rf build && mkdir build && cd build && cmake -DGGML_CUBLAS=1 .. && cmake --build . --config Release

Compiling on Windows: developer cmp-nct notes: 'I personally compile it using VScode. When compiling with CUDA support using the Microsoft compiler it's essential to select the "Community edition build tools". Otherwise CUDA won't compile.'

Once compiled you can then use bin/falcon_main just like you would use llama.cpp. For example:

bin/falcon_main -t 8 -ngl 100 -b 1 -m falcon7b-instruct.ggmlv3.q4_0.bin -p "What is a falcon?\n### Response:"

You can specify -ngl 100 regardles of your VRAM, as it will automatically detect how much VRAM is available to be used.

Adjust -t 8 (the number of CPU cores to use) according to what performs best on your system. Do not exceed the number of physical CPU cores you have.

-b 1 reduces batch size to 1. This slightly lowers prompt evaluation time, but frees up VRAM to load more of the model on to your GPU. If you find prompt evaluation too slow and have enough spare VRAM, you can remove this parameter.

Please see https://github.com/cmp-nct/ggllm.cpp for further details and instructions.

Provided files

Name	Quant method	Bits	Size	Max RAM required	Use case
falcon-40b-sft-mix-1226.ggccv1.q2_K.bin	q2_K	2	13.74 GB	16.24 GB	New k-quant method. Uses GGML_TYPE_Q4_K for the attention.vw and feed_forward.w2 tensors, GGML_TYPE_Q2_K for the other tensors.
falcon-40b-sft-mix-1226.ggccv1.q3_K.bin	q3_K_S	3	17.98 GB	20.48 GB	New k-quant method. Uses GGML_TYPE_Q3_K for all tensors
falcon-40b-sft-mix-1226.ggccv1.q4_K.bin	q4_K_S	4	23.54 GB	26.04 GB	New k-quant method. Uses GGML_TYPE_Q4_K for all tensors
falcon-40b-sft-mix-1226.ggccv1.q5_K.bin	q5_K_S	5	28.77 GB	31.27 GB	New k-quant method. Uses GGML_TYPE_Q5_K for all tensors
falcon-40b-sft-mix-1226.ggccv1.q6_K.bin	q6_K	6	34.33 GB	36.83 GB	New k-quant method. Uses GGML_TYPE_Q8_K - 6-bit quantization - for all tensors
falcon-40b-sft-mix-1226.ggccv1.q8_0.bin	q8_0	8	44.46 GB	46.96 GB	Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users.

Note: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.

Discord

For further support, and discussions on these models and AI in general, join us at:

TheBloke AI's Discord server

Thanks, and how to contribute.

Thanks to the chirper.ai team!

I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.

If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.

Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.

• Patreon: https://patreon.com/TheBlokeAI
• Ko-Fi: https://ko-fi.com/TheBlokeAI

Special thanks to: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.

Patreon special mentions: Spiking Neurons AB, Kevin Schuppel, Cory Kujawski, senxiiz, Luke Pendergrass, John Villwock, Ghost , Alex , Sean Connelly, Space Cruiser, Eugene Pentland, Pyrater, Matthew Berman, Dave, Derek Yates, Jonathan Leane, Viktor Bowallius, Michael Levine, Joseph William Delisle, Fred von Graf, Asp the Wyvern, Nikolai Manek, Pierre Kircher, webtim, K, RoA, Karl Bernard, Artur Olbinski, Rainer Wilmers, Ai Maven, Nathan LeClaire, Ajan Kanaga, Stephen Murray, Edmond Seymore, zynix , Imad Khwaja, John Detwiler, Randy H, subjectnull, Alps Aficionado, Greatston Gnanesh, Trenton Dambrowitz, Junyu Yang, Raven Klaugh, biorpg, Deep Realms, vamX, Talal Aujan, Johann-Peter Hartmann, WelcomeToTheClub, Chris McCloskey, Luke, chris gileta, terasurfer , Iucharbius , Preetika Verma, Willem Michiel, Fen Risland, SuperWojo, Khalefa Al-Ahmad, Daniel P. Andersen, Gabriel Puliatti, Illia Dulskyi, Willian Hasse, Oscar Rangel, ya boyyy, Mano Prime, Lone Striker, Kalila

Thank you to all my generous patrons and donaters!

Original model card: Open Assistant's Falcon 40B SFT MIX

Open-Assistant Falcon 40B SFT MIX Model

This model is a fine-tuning of TII's Falcon 40B LLM. It was trained on a mixture of OASST top-2 threads (exported on June 2, 2023), Dolly-15k and synthetic instruction datasets (see dataset configuration below).

Model Details

• Finetuned from: tiiuae/falcon-40b
• Model type: Causal decoder-only transformer language model
• Language: English, German, Spanish, French (and limited capabilities in Italian, Portuguese, Polish, Dutch, Romanian, Czech, Swedish);
• Demo: Continuations for 250 random prompts, multiligual-60
• Eval results: ilm-eval
• Weights & Biases: Training log (checkpoint: 1226 steps)
• License: Apache 2.0
• Contact: Open-Assistant Discord

Prompting

Two special tokens are used to mark the beginning of user and assistant turns: <|prompter|> and <|assistant|>. Each turn ends with a <|endoftext|> token.

Input prompt example:

<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>

The input ends with the <|assistant|> token to signal that the model should start generating the assistant reply.

Configuration Details

Model:

falcon-40b:
  dtype: bf16
  learning_rate: 1e-5
  model_name: "tiiuae/falcon-40b"
  deepspeed_config: configs/zero3_config_falcon.json
  weight_decay: 0.0
  max_length: 2048
  warmup_steps: 20
  gradient_checkpointing: true
  gradient_accumulation_steps: 1
  per_device_train_batch_size: 18
  per_device_eval_batch_size: 10
  eval_steps: 120
  save_strategy: steps
  save_steps: 613
  num_train_epochs: 8
  save_total_limit: 4
  use_flash_attention: false
  residual_dropout: 0.3
  residual_dropout_lima: true

Dataset:

sft9-stage2:
  # oasst_export: 100.00% (29899)
  # vicuna: 50.00% (16963)
  # code_alpaca: 50.00% (9510)
  # oa_wiki_qa_bart_10000row: 100.00% (9434)
  # grade_school_math_instructions: 100.00% (8351)
  # dolly15k: 100.00% (14250)

  use_custom_sampler: true
  datasets:
    - oasst_export:
        lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk" # sft-8.0
        input_file_path: 2023-06-02_oasst_all_labels.jsonl.gz
        val_split: 0.05
        top_k: 2
    - vicuna:
        fraction: 0.5
        val_split: 0.025
        max_val_set: 250
    - code_alpaca:
        fraction: 0.5
        val_split: 0.05
        max_val_set: 250
    - oa_wiki_qa_bart_10000row:
        val_split: 0.05
        max_val_set: 250
    - grade_school_math_instructions:
        val_split: 0.05
    - dolly15k:
        val_split: 0.05
        max_val_set: 300