model-index:
- name: notus-7b-v1
results: []
datasets:
- argilla/ultrafeedback-binarized-preferences
language:
- en
base_model: alignment-handbook/zephyr-7b-sft-full
library_name: transformers
pipeline_tag: text-generation
tags:
- dpo
- preference
- ultrafeedback
license: mit
Model Card for Notus 7B v1
Notus is a collection of fine-tuned models using Direct Preference Optimization (DPO) and related RLHF techniques. This model is version 1, fine-tuned with DPO starting with zephyr-7b-beta's SFT model.
Following a data-first approach, the only difference between Notus-7B-v1 and Zephyr-7B-beta is the preference dataset used for dDPO. In particular, we've found data issues in the original UltraFeedback dataset, leading to high-scores for bad responses. After curating several hundreds of data points, we decided to binarize the dataset using the preference ratings, instead of the original critique overall_score.
Using preference ratings, instead of critiques scores, led to a new dataset where the chosen response is different in ~50% of the cases.
This model wouldn't have been possible without the amazing Alignment Handbook and it's based on fruitful discussions with the H4 team. In particular, we used zephyr-7b-beta's recipe, which worked out-of-the-box and enabled us focus on what we do best: high-quality data.
Notus models are intended to be used as assistants via chat-like applications, and are evaluated with Chat (MT-Bench, AlpacaEval) and Academic (Open LLM Leaderboard) benchmarks for a direct comparison with the original Zephyr dDPO model and other 7B models.
Model Details
Model Description
- Developed by: Argilla (based on HuggingFace H4 and MistralAI previous efforts and amazing work)
- Shared by: Argilla
- Model type: GPT-like 7B model DPO fine-tuned
- Language(s) (NLP): Mainly English
- License: MIT (same as Zephyr 7B-beta)
- Finetuned from model:
alignment-handbook/zephyr-7b-sft-full
Model Sources
- Repository: https://github.com/argilla-io/notus-7b
- Paper: N/A
- Demo: https://argilla-notus-chat-ui.hf.space/
Performance
Chat benchmarks
Table adapted from Zephyr-7b-β and Starling's original tables for MT-Bench and AlpacaEval benchmarks. Results are shown sorted by AlpacaEval win rates and ommit some >7B for brevity. Notus stays on par with Zephyr on MT-Bench, while surpassing Zephyr, Claude 2, and Cohere Command on AlpacaEval. Making Notus the most-competitive 7B commercial model on AlpacaEval.
| Model | Size | Alignment | MT-Bench (score) | AlpacaEval (win rate %) | License |
|---|---|---|---|---|---|
| GPT-4-turbo | - | ? | 9.32 | 97.70 | Proprietary |
| XwinLM 70b V0.1 | 70B | dPPO | - | 95.57 | LLaMA 2 License |
| GPT-4 | - | RLHF | 8.99 | 95.03 | Proprietary |
| Tulu 2+DPO 70B V0.1 | 70B | dDPO | 6.29 | 95.28 | Proprietary |
| LLaMA2 Chat 70B | 70B | RLHF | 6.86 | 92.66 | LLaMA 2 License |
| Starling-7B | 7B | C-RLFT + APA | 8.09 | 91.99 | CC-BY-NC-4.0 |
| Notus-7b-v1 | 7B | dDPO | 7.30 | 91.42 | MIT |
| Claude 2 | - | RLHF | 8.06 | 91.36 | Proprietary |
| Zephyr-7b-β | 7B | dDPO | 7.34 | 90.60 | MIT |
| Cohere Command | - | RLHF | - | 90.62 | Proprietary |
| GPT-3.5-turbo | - | RLHF | 7.94 | 89.37 | Proprietary |
Academic benchmarks
Results from OpenLLM Leaderboard:
| Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K | DROP |
|---|---|---|---|---|---|---|---|---|
| Zephyr 7B dDPO (HuggingFaceH4/zephyr-7b-beta) | 52.15 | 62.03 | 84.36 | 61.07 | 57.45 | 77.74 | 12.74 | 9.66 |
| argilla/notus-7b-v1 | 52.89 | 64.59 | 84.78 | 63.03 | 54.37 | 79.4 | 15.16 | 8.91 |
Training Details
Training Hardware
We used a VM with 8 x A100 40GB hosted in Lambda Labs.
Training Data
We used a a new curated version of openbmb/UltraFeedback, named argilla/ultrafeedback-binarized-avg-rating-for-dpo.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-07
- train_batch_size: 8
- eval_batch_size: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 64
- total_eval_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 3
Training results
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.5051 | 0.1 | 100 | 0.5180 | 0.1475 | -0.3954 | 0.7183 | 0.5429 | -246.6286 | -297.5412 | -2.7438 | -3.0431 |
| 0.4321 | 0.21 | 200 | 0.4375 | 0.1353 | -0.9529 | 0.7540 | 1.0882 | -252.2036 | -297.6632 | -2.7578 | -3.0543 |
| 0.3848 | 0.31 | 300 | 0.4301 | -0.4813 | -1.8921 | 0.7302 | 1.4107 | -261.5956 | -303.8301 | -2.7592 | -3.0508 |
| 0.3777 | 0.42 | 400 | 0.4091 | -0.8597 | -2.5306 | 0.7698 | 1.6709 | -267.9805 | -307.6138 | -2.7476 | -3.0474 |
| 0.3559 | 0.52 | 500 | 0.4332 | -1.0424 | -2.6019 | 0.7619 | 1.5595 | -268.6939 | -309.4406 | -2.2960 | -2.6106 |
| 0.4178 | 0.62 | 600 | 0.3934 | -0.6434 | -2.4837 | 0.7659 | 1.8404 | -267.5121 | -305.4503 | -2.5487 | -2.8508 |
| 0.4206 | 0.73 | 700 | 0.4058 | -1.4700 | -3.5113 | 0.7857 | 2.0413 | -277.7877 | -313.7168 | -2.5679 | -2.8727 |
| 0.4323 | 0.83 | 800 | 0.3929 | -0.9025 | -2.6935 | 0.7897 | 1.7910 | -269.6095 | -308.0414 | -2.6213 | -2.9202 |
| 0.3706 | 0.93 | 900 | 0.3903 | -1.1122 | -3.0257 | 0.8056 | 1.9135 | -272.9316 | -310.1388 | -2.5428 | -2.8416 |
| 0.0496 | 1.04 | 1000 | 0.3991 | -1.4248 | -4.1245 | 0.8016 | 2.6997 | -283.9196 | -313.2651 | -2.5093 | -2.8150 |
| 0.0723 | 1.14 | 1100 | 0.3999 | -1.8789 | -4.5317 | 0.7897 | 2.6528 | -287.9914 | -317.8056 | -2.5170 | -2.8242 |
| 0.0481 | 1.25 | 1200 | 0.4191 | -2.6211 | -5.5294 | 0.7817 | 2.9083 | -297.9687 | -325.2281 | -2.5139 | -2.8109 |
| 0.0432 | 1.35 | 1300 | 0.4070 | -2.0605 | -5.0460 | 0.8056 | 2.9855 | -293.1345 | -319.6214 | -2.5153 | -2.8121 |
| 0.0402 | 1.45 | 1400 | 0.4001 | -2.2445 | -5.0942 | 0.7937 | 2.8497 | -293.6164 | -321.4614 | -2.4383 | -2.7388 |
| 0.0529 | 1.56 | 1500 | 0.4066 | -2.3499 | -5.2468 | 0.8016 | 2.8969 | -295.1426 | -322.5153 | -2.3906 | -2.6963 |
| 0.0651 | 1.66 | 1600 | 0.3962 | -2.0597 | -4.8915 | 0.8016 | 2.8318 | -291.5901 | -319.6136 | -2.3390 | -2.6469 |
| 0.0738 | 1.77 | 1700 | 0.3942 | -1.8893 | -4.6107 | 0.8135 | 2.7214 | -288.7817 | -317.9099 | -2.3532 | -2.6607 |
| 0.0597 | 1.87 | 1800 | 0.3990 | -1.8774 | -4.7221 | 0.8175 | 2.8448 | -289.8961 | -317.7905 | -2.2728 | -2.5908 |
| 0.0686 | 1.97 | 1900 | 0.3924 | -1.8745 | -4.6807 | 0.8056 | 2.8062 | -289.4821 | -317.7617 | -2.2554 | -2.5658 |
| 0.0116 | 2.08 | 2000 | 0.4260 | -2.4687 | -5.7190 | 0.7937 | 3.2503 | -299.8647 | -323.7037 | -2.2297 | -2.5347 |
| 0.0114 | 2.18 | 2100 | 0.4519 | -2.8266 | -6.3706 | 0.7976 | 3.5440 | -306.3802 | -327.2823 | -2.2185 | -2.5219 |
| 0.0073 | 2.28 | 2200 | 0.4563 | -2.9422 | -6.5564 | 0.8016 | 3.6142 | -308.2384 | -328.4384 | -2.2103 | -2.5126 |
| 0.0094 | 2.39 | 2300 | 0.4636 | -3.3246 | -7.0542 | 0.8016 | 3.7296 | -313.2165 | -332.2628 | -2.2059 | -2.5081 |
| 0.0056 | 2.49 | 2400 | 0.4745 | -3.3599 | -7.1652 | 0.7976 | 3.8053 | -314.3266 | -332.6161 | -2.1945 | -2.4943 |
| 0.0052 | 2.6 | 2500 | 0.4812 | -3.4916 | -7.3391 | 0.7976 | 3.8475 | -316.0656 | -333.9322 | -2.1888 | -2.4881 |
| 0.0065 | 2.7 | 2600 | 0.4678 | -3.2226 | -6.9887 | 0.7976 | 3.7661 | -312.5613 | -331.2425 | -2.1644 | -2.4560 |
| 0.0059 | 2.8 | 2700 | 0.4694 | -3.4307 | -7.2484 | 0.7976 | 3.8177 | -315.1584 | -333.3234 | -2.1572 | -2.4483 |
| 0.0054 | 2.91 | 2800 | 0.4707 | -3.4959 | -7.3283 | 0.8056 | 3.8324 | -315.9576 | -333.9758 | -2.1575 | -2.4491 |
Framework versions
- Transformers 4.35.0
- Pytorch 2.1.1+cu121
- Datasets 2.14.6
- Tokenizers 0.14.1
Evaluation during Training
- Loss: 0.4730
- Rewards/chosen: -3.5289
- Rewards/rejected: -7.3700
- Rewards/accuracies: 0.8016
- Rewards/margins: 3.8412
- Logps/rejected: -316.3751
- Logps/chosen: -334.3053
- Logits/rejected: -2.1644
- Logits/chosen: -2.4556