library_name: peft
base_model: mistralai/Mistral-7B-v0.1
license: apache-2.0
datasets:
- upaya07/NeurIPS-LLM-data
language:
- en
tags:
- NeurIPS
- NeurIPS LLM Efficiency Challenge
- NeurIPS LLM Efficiency Challenge Winner Model
- Team Upaya
Model Card for Model ID
🚀🚀🚀 Our model Birbal-7B-V1 achieved 🏆 first rank 🏆 in among 80+ global teams in NeurIPS Large Language Model Efficiency Challenge: 1 LLM + 1GPU + 1Day organized by Microsoft and Meta.
📣 P.S.: Please reach out to us, if you would be interested in supporting compute resources. Here are our recent achievements in LLM space: https://upaya.ai/
Model Details
Birbal-7B-V1 is fine-tuned on our curated dataset of 200k size for nearly 3 epochs. Our approach for dataset preparation is focused on finding most-relavant examples from large pool of tasks spanning across NLP, Maths, Commonsense, etc. Hence, we expect model to perform well on different tasks including unseen tasks.
Model Description
- Project GitHub Page: https://github.com/Upaya07/NeurIPS-llm-efficiency-challenge
- Developed by: ❤️ Team Upaya - Ashvini Kumar Jindal, Ankur Parikh, Pawan Rajpoot
- Funded by: self-work
- Model type: fine-tuned. It is a PEFT model and can be combined with Mistral-7B model.
- Language(s) (NLP): English
- License: Apache-2.0
- Finetuned from model: mistralai/Mistral-7B-v0.1
Model Sources [optional]
Uses
Birbal-7B-V1 is trained with the following format:
##Instruction
<instruction>
##Input
<input>
##Response
<response>
If a record does not contain any instruction, here is the training format:
##Input
<input>
##Response
<response>
It will performed best if queried in the same way.
Downstream Use
Birbal-7B-V1 is fine-tuned on our curated dataset that contain examples from large number of tasks spanning across NLP, Maths, QA, etc. Hence, we expect the model to perform well on in general on various kinds of tasks.
How to Get Started with the Model
It is quite easy! Merge Birbal-7B-V1 peft model with Mistral-7B model and start running inference!
Training Details
We used Mistral-7B as a base model and fine-tuned it on a single RTX 4090 GPU for 24 hours as per the competition rules. Fine-tuning was performed using 4-bit QLoRA.
Training Data
Here is high-level diagram of our data preparation strategy:
Please visit https://huggingface.co/datasets/upaya07/NeurIPS-LLM-data for more details.
Training Hyperparameters
Refer to https://github.com/Upaya07/NeurIPS-llm-efficiency-challenge/blob/main/training/axolotl/examples/mistral/nips/nips_02.yml for example set of hyperparams used.
Evaluation
Results
| Task | Score |
|---|---|
| MMLU - EM | 0.629 |
| MMLU - EM (Robustness) | 0.591 |
| MMLU - EM (Fairness) | 0.596 |
| MMLU Mean Win Rate | 0.417 |
| TruthfulQA - EM | 0.59 |
| TruthfulQA - EM (Robustness) | 0.541 |
| TruthfulQA - EM (Fairness) | 0.492 |
| TruthfulQA Mean Win Rate | 0.75 |
| BIG-bench - EM | 0.330 |
| BIG-bench Mean Win Rate | 0.75 |
| GSM8K - EM | 0.443 |
| GSM8K Mean Win Rate | 0.625 |
| BBQ - EM | 0.738 |
| BBQ Mean Win Rate | 0.25 |
| sam_sum - ROUGE-2 | 0.127 |
| sam_sum - Stereotypes (race) | 0.667 |
| sam_sum - Stereotypes (gender) | 0.447 |
| sam_sum - Representation (race) | 0.458 |
| sam_sum - Representation (gender) | 0.013 |
| sam_sum Mean Win Rate | 0.383 |
| corr2cause - EM | 0.615 |
| corr2cause Mean Win Rate | 0.875 |
| MATH (chain-of-thoughts) - Equivalent (chain of thought) | 0.121 |
| MATH Mean Win Rate | 0.75 |
| ethics_justice - EM | 0.68 |
| ethics_justice - EM (Robustness) | 0.645 |
| ethics_justice - EM (Fairness) | 0.62 |
| ethics_commonsense - EM | 0.41 |
| ethics_commonsense - EM (Robustness) | 0.33 |
| ethics_commonsense - EM (Fairness) | 0.345 |
| ethics_virtue - EM | 0.895 |
| ethics_virtue - EM (Robustness) | 0.865 |
| ethics_virtue - EM (Fairness) | 0.86 |
| ethics_deontology - EM | 0.63 |
| ethics_deontology - EM (Robustness) | 0.585 |
| ethics_deontology - EM (Fairness) | 0.595 |
| ethics_utilitarianism - EM | 0.72 |
| ethics_utilitarianism - EM (Robustness) | 0.6 |
| ethics_utilitarianism - EM (Fairness) | 0.645 |
| ethics Mean Win Rate | 0.55 |
| 🔥 Score_full | 0.579 |
| 🔥 Score_open | 0.516 |
| 🔥 Score_hidden | 0.61 |
Top-5 Teams
| Position | Score |
|---|---|
| 5th rank | 0.362 |
| 4th rank | 0.371 |
| 3rd rank | 0.381 |
| 2nd rank | 0.424 |
| 🔥 Ours (1st) | 0.579 |
Citation [optional]
BibTeX:
[More Information Needed]
APA:
[More Information Needed]
Glossary [optional]
[More Information Needed]
More Information [optional]
[More Information Needed]
Model Card Authors [optional]
[More Information Needed]
Model Card Contact
[More Information Needed]
Training procedure
The following bitsandbytes quantization config was used during training:
- quant_method: bitsandbytes
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16
Framework versions
- PEFT 0.6.1