DavidGF's picture
Update README.md
f9d5575 verified
metadata
license: other
license_name: gemma-terms-of-use
license_link: https://ai.google.dev/gemma/terms
language:
  - de
  - en
tags:
  - sft
  - laserRMT
  - laser-QLoRa
  - finetune
  - work in progress
  - alpha

SauerkrautLM

VAGO solutions SauerkrautLM-Gemma-2b (alpha)

Introducing SauerkrautLM-Gemma-2b – our German Sauerkraut version of the powerful google/gemma-2b ! It is an early stage finetuned model and should be used with caution!

The model SauerkrautLM-Gemma-2b is a joint effort between VAGO solutions and Hyperspace.ai. Much appreciation goes to the tremendous research effort of Fernando Fernandes Neto, David Golchinfar and Eric Hartford on their laserRMT approach. Without their independent research collaboration this model release would not have been possible.

  • Fintuned with SFT
  • Using a novel training technique: laser-QLoRA - we partially freeze the model according to a laser-like analysis (Official Paper soon). It allows to evaluate the no free lunch theorem and supports better decision making when optimizing the theorem - created by the LaserRMT research group
  • Optimized with LaserRMT

Table of Contents

  1. Overview of all SauerkrautLM-Gemma-2b models
  2. Model Details
  3. Evaluation
  4. Disclaimer
  5. Contact
  6. Collaborations
  7. Acknowledgement

All SauerkrautLM-Gemma-2b Models

Model HF GPTQ GGUF AWQ
SauerkrautLM-Gemma-2b Link coming soon coming soon coming soon

Model Details

SauerkrautLM-Gemma-2b

Training procedure:

Warning: This finetuned model is in an early stage and we sometimes observed strange behavior. It is still work in progress!

Anyone who has attempted or succeeded in fine-tuning a model is aware of the difficulty in nudging it towards a specific skill, such as mastering new languages, as well as the challenges associated with achieving significant improvements in performance. Experimenting with a novel training strategy and Spherical Linear Interpolation alongside a lasered version of the model itself has proven to be both fascinating and revealing.

Furthermore, we developed one iteration of the model using our entire SFT -Sauerkraut dataset and two additional iterations using subsets of the full dataset—one focused on enhancing MMLU and TQA capabilities, and the other on boosting GSM8K and Winogrande skills.

We actively monitor and assesed the results of each training. Whenever we found a decrease in perplexity on the gsm8k benchmark we intervined. By following this procedure we were able to improve the overall performance, especially in math abilities, without detracting from performance on other benchmarks—a task that is, in general, quite difficult.

This process not only helps in understanding the effectiveness of Spherical Linear Interpolation but also introduces a new method for refining models with enhanced skills through a cycle of targeted data selection (Laser data(x)) + SLERP, followed by a subsequent focus on different data (Laser again on data(y)).

Additionally, we integrated a novel training strategy on the SFT training process, where we partially freeze the model according to a laser-like analysis aiming to navigate and optimize the trade-offs highlighted by the no free lunch theorem. This innovative training method effectively prevents the significant problem of language models forgetting previously acquired knowledge. This aspect is particularly crucial when attempting to teach the model specific skills, such as a new language, where in general, the model might lose a considerable amount of its prior knowledge and exhibit a decline in overall intelligence.

Detailed information on how the new training strategy works and the advantages it offers over conventional training methods will soon be published in a detailed paper by the LaserRMT research group.

We teached German language skills on this model. As far as we know, it is the first Gemma-2b model with bilingual skills in German and English. Nevertheless, formulations may occur that are not entirely correct (still work in progress).

Prompt Template:

We trained on vicuna prompt template. Please add the following stopping string to your client: "</s>","</p>" (we did not add the special tokens to the training config)

You are a helpful AI Assistant.

USER: Hello, how are you?
ASSISTANT:

Evaluation

(with lm-evaluation-harness 0.4.1)

Open LLM Leaderboard:

Metric Value
Avg. 48.93
ARC (25-shot) 49.32
HellaSwag (10-shot) 71.23
MMLU (5-shot) 42.06
TruthfulQA (0-shot) 35.73
Winogrande (5-shot) 67.56
GSM8K (5-shot) 27.67

Performance

Model AGIEval GPT4All TruthfulQA BigBench Average ⬇️
VAGOsolutions/SauerkrautLM-Gemma-7b 37.5 72.46 61.24 45.33 54.13
zephyr-7b-beta 37.52 71.77 55.26 39.77 51.08
zephyr-7b-gemma-v0.1 34.22 66.37 52.19 37.10 47.47
VAGOsolutions/SauerkrautLM-Gemma-2b 24.28 63.59 35.73 22.77 36.59
google/gemma-7b-it 21.33 40.84 41.70 30.25 33.53
Details of AGIEval, GPT4All, TruthfulQA, BigBench

AGIEval

Tasks Version Filter n-shot Metric Value Stderr
agieval_sat_math 1 none None acc 0.2409 ± 0.0289
none None acc_norm 0.2455 ± 0.0291
agieval_sat_en_without_passage 1 none None acc 0.3010 ± 0.0320
none None acc_norm 0.2816 ± 0.0314
agieval_sat_en 1 none None acc 0.3301 ± 0.0328
none None acc_norm 0.2961 ± 0.0319
agieval_lsat_rc 1 none None acc 0.2007 ± 0.0245
none None acc_norm 0.1933 ± 0.0241
agieval_lsat_lr 1 none None acc 0.1941 ± 0.0175
none None acc_norm 0.2039 ± 0.0179
agieval_lsat_ar 1 none None acc 0.2304 ± 0.0278
none None acc_norm 0.2391 ± 0.0282
agieval_logiqa_en 1 none None acc 0.2089 ± 0.0159
none None acc_norm 0.2581 ± 0.0172
agieval_aqua_rat 1 none None acc 0.2480 ± 0.0272
none None acc_norm 0.2244 ± 0.0262

Average: 24.28%

GPT4All

Tasks Version Filter n-shot Metric Value Stderr
arc_challenge 1 none None acc 0.4334 ± 0.0145
none None acc_norm 0.4309 ± 0.0145
arc_easy 1 none None acc 0.7433 ± 0.0090
none None acc_norm 0.7264 ± 0.0091
boolq 2 none None acc 0.7165 ± 0.0079
hellaswag 1 none None acc 0.5357 ± 0.0050
none None acc_norm 0.7158 ± 0.0045
openbookqa 1 none None acc 0.318 ± 0.0208
none None acc_norm 0.402 ± 0.0219
piqa 1 none None acc 0.7709 ± 0.0098
none None acc_norm 0.7807 ± 0.0097
winogrande 1 none None acc 0.6788 ± 0.0131

Average: 63.59%

TruthfulQA

Tasks Version Filter n-shot Metric Value Stderr
truthfulqa_mc2 2 none 0 acc 0.3573 ± 0.0135

Average: 35.73%

Bigbench

Tasks Version Filter n-shot Metric Value Stderr
bbh_zeroshot_tracking_shuffled_objects_three_objects 2 flexible-extract 0 exact_match 0.3280 ± 0.0298
bbh_zeroshot_tracking_shuffled_objects_seven_objects 2 flexible-extract 0 exact_match 0.1120 ± 0.0200
bbh_zeroshot_tracking_shuffled_objects_five_objects 2 flexible-extract 0 exact_match 0.1520 ± 0.0228
bbh_zeroshot_temporal_sequences 2 flexible-extract 0 exact_match 0.1000 ± 0.0190
bbh_zeroshot_sports_understanding 2 flexible-extract 0 exact_match 0.5360 ± 0.0316
bbh_zeroshot_snarks 2 flexible-extract 0 exact_match 0.2753 ± 0.0336
bbh_zeroshot_salient_translation_error_detection 2 flexible-extract 0 exact_match 0.1400 ± 0.0220
bbh_zeroshot_ruin_names 2 flexible-extract 0 exact_match 0.1120 ± 0.0200
bbh_zeroshot_reasoning_about_colored_objects 2 flexible-extract 0 exact_match 0.1080 ± 0.0197
bbh_zeroshot_navigate 2 flexible-extract 0 exact_match 0.5800 ± 0.0313
bbh_zeroshot_movie_recommendation 2 flexible-extract 0 exact_match 0.4360 ± 0.0314
bbh_zeroshot_logical_deduction_three_objects 2 flexible-extract 0 exact_match 0.0000 ± 0.0000
bbh_zeroshot_logical_deduction_seven_objects 2 flexible-extract 0 exact_match 0.0720 ± 0.0164
bbh_zeroshot_logical_deduction_five_objects 2 flexible-extract 0 exact_match 0.0000 ± 0.0000
bbh_zeroshot_geometric_shapes 2 flexible-extract 0 exact_match 0.0000 ± 0.0000
bbh_zeroshot_disambiguation_qa 2 flexible-extract 0 exact_match 0.3400 ± 0.0300
bbh_zeroshot_date_understanding 2 flexible-extract 0 exact_match 0.3360 ± 0.0299
bbh_zeroshot_causal_judgement 2 flexible-extract 0 exact_match 0.4706 ± 0.0366

Average: 22.77%

Disclaimer

We must inform users that despite our best efforts in data cleansing, the possibility of uncensored content slipping through cannot be entirely ruled out. However, we cannot guarantee consistently appropriate behavior. Therefore, if you encounter any issues or come across inappropriate content, we kindly request that you inform us through the contact information provided. Additionally, it is essential to understand that the licensing of these models does not constitute legal advice. We are not held responsible for the actions of third parties who utilize our models.  

Contact

If you are interested in customized LLMs for business applications, please get in contact with us via our websites. We are also grateful for your feedback and suggestions.  

Collaborations

We are also keenly seeking support and investment for our startups, VAGO solutions and Hyperspace where we continuously advance the development of robust language models designed to address a diverse range of purposes and requirements. If the prospect of collaboratively navigating future challenges excites you, we warmly invite you to reach out to us at VAGO solutions, Hyperspace.computer

Acknowledgement

Many thanks to google for providing such valuable model to the Open-Source community