Edit model card

Update

  • 01.03.2024 - Reuploaded the model in bfloat16 dtype.
  • 02.03.2024 - strongest Gemma finetune model so far: added AGIEval,GPT4ALL and Bigbench scoring with AVG of 54.13 and H6 Open LLM Leaderboard with AVG of 67.83

SauerkrautLM

VAGO solutions SauerkrautLM-Gemma-7b (alpha)

Introducing SauerkrautLM-Gemma-7b – our Sauerkraut version of the powerful google/gemma-7b ! It is an early stage finetuned model and should be used with caution!

The model SauerkrautLM-Gemma-7b 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
  • Aligned with DPO
  • 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-7b models
  2. Model Details
  3. Evaluation
  4. Disclaimer
  5. Contact
  6. Collaborations
  7. Acknowledgement

All SauerkrautLM-Gemma-7b Models

Model HF GPTQ GGUF AWQ
SauerkrautLM-Gemma-7b Link coming soon coming soon coming soon

Model Details

SauerkrautLM-Gemma-7b

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.

After optimizing our primary SFT model, we applied a similar strategy to our new DPO Dataset, dividing it into further subsets. We trained one model on the entire dataset again and two more on these specialized subsets.

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 and DPO 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 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

Open LLM Leaderboard:

Metric Value
Avg. 67.83
ARC (25-shot) 59.98
HellaSwag (10-shot) 81.91
MMLU (5-shot) 63.76
TruthfulQA (0-shot) 61
Winogrande (5-shot) 76.64
GSM8K (5-shot) 63.68

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
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 0.3682 ± 0.0326
none None acc_norm 0.3364 ± 0.0319
agieval_sat_en_without_passage 1 none None acc 0.4272 ± 0.0345
none None acc_norm 0.3738 ± 0.0338
agieval_sat_en 1 none None acc 0.7427 ± 0.0305
none None acc_norm 0.6893 ± 0.0323
agieval_lsat_rc 1 none None acc 0.5539 ± 0.0304
none None acc_norm 0.5167 ± 0.0305
agieval_lsat_lr 1 none None acc 0.3431 ± 0.0210
none None acc_norm 0.3471 ± 0.0211
agieval_lsat_ar 1 none None acc 0.1913 ± 0.0260
none None acc_norm 0.1739 ± 0.0250
agieval_logiqa_en 1 none None acc 0.3303 ± 0.0184
none None acc_norm 0.3303 ± 0.0184
agieval_aqua_rat 1 none None acc 0.2480 ± 0.0272
none None acc_norm 0.2323 ± 0.0265

Average: 37.5%

GPT4All

Tasks Version Filter n-shot Metric Value Stderr
arc_challenge 1 none None acc 0.5358 ± 0.0146
none None acc_norm 0.5597 ± 0.0145
arc_easy 1 none None acc 0.8249 ± 0.0078
none None acc_norm 0.7955 ± 0.0083
boolq 2 none None acc 0.8651 ± 0.006
hellaswag 1 none None acc 0.6162 ± 0.0049
none None acc_norm 0.8117 ± 0.0039
openbookqa 1 none None acc 0.336 ± 0.0211
none None acc_norm 0.470 ± 0.0223
piqa 1 none None acc 0.7900 ± 0.0095
none None acc_norm 0.8096 ± 0.00
winogrande 1 none None acc 0.7609 ± 0.012

Average: 72.46%

TruthfulQA

Tasks Version Filter n-shot Metric Value Stderr
truthfulqa_mc2 2 none 0 acc 0.6124 ± 0.0148

Average: 61.24%

Bigbench

Tasks Version Filter n-shot Metric Value Stderr
bbh_zeroshot_tracking_shuffled_objects_three_objects 2 flexible-extract 0 exact_match 0.2760 ± 0.0283
bbh_zeroshot_tracking_shuffled_objects_seven_objects 2 flexible-extract 0 exact_match 0.1280 ± 0.0212
bbh_zeroshot_tracking_shuffled_objects_five_objects 2 flexible-extract 0 exact_match 0.1240 ± 0.0209
bbh_zeroshot_temporal_sequences 2 flexible-extract 0 exact_match 0.4520 ± 0.0315
bbh_zeroshot_sports_understanding 2 flexible-extract 0 exact_match 0.7120 ± 0.0287
bbh_zeroshot_snarks 2 flexible-extract 0 exact_match 0.5056 ± 0.0376
bbh_zeroshot_salient_translation_error_detection 2 flexible-extract 0 exact_match 0.4480 ± 0.0315
bbh_zeroshot_ruin_names 2 flexible-extract 0 exact_match 0.4520 ± 0.0315
bbh_zeroshot_reasoning_about_colored_objects 2 flexible-extract 0 exact_match 0.4800 ± 0.0317
bbh_zeroshot_navigate 2 flexible-extract 0 exact_match 0.5480 ± 0.0315
bbh_zeroshot_movie_recommendation 2 flexible-extract 0 exact_match 0.7000 ± 0.0290
bbh_zeroshot_logical_deduction_three_objects 2 flexible-extract 0 exact_match 0.5200 ± 0.0317
bbh_zeroshot_logical_deduction_seven_objects 2 flexible-extract 0 exact_match 0.4120 ± 0.0312
bbh_zeroshot_logical_deduction_five_objects 2 flexible-extract 0 exact_match 0.3840 ± 0.0308
bbh_zeroshot_geometric_shapes 2 flexible-extract 0 exact_match 0.2920 ± 0.0288
bbh_zeroshot_disambiguation_qa 2 flexible-extract 0 exact_match 0.6480 ± 0.0303
bbh_zeroshot_date_understanding 2 flexible-extract 0 exact_match 0.5000 ± 0.0317
bbh_zeroshot_causal_judgement 2 flexible-extract 0 exact_match 0.5775 ± 0.0362

Average: 45.33%

Despite the fact that we archived great results on the Open LLM leaderboard benchmarks the model subjectively does not feel as smart as comparable mistral finetunes. Most of its answers are coherent but we observed that the model sometimes answers realy lazy or odd.

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

Downloads last month
6,829
Safetensors
Model size
8.54B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for VAGOsolutions/SauerkrautLM-Gemma-7b

Merges
1 model
Quantizations
2 models

Space using VAGOsolutions/SauerkrautLM-Gemma-7b 1

Collection including VAGOsolutions/SauerkrautLM-Gemma-7b