Downloads · 30 days
26
7% of all-time downloads
RichardErkhov/argilla_-_notux-8x7b-v1-gguf
argilla_-_notux-8x7b-v1-gguf is a machine learning model from RichardErkhov. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
26
7% of all-time downloads
All-time downloads
350
Public
Repo size
530 GB
Likes
0
Public
Click a slice to open those files.
.gguf614 GB · 100%
From the Hugging Face model README
Quantization made by Richard Erkhov.
notux-8x7b-v1 - GGUF
| Name | Quant method | Size |
|---|---|---|
| notux-8x7b-v1.Q2_K.gguf | Q2_K | 16.12GB |
| notux-8x7b-v1.IQ3_XS.gguf | IQ3_XS | 18.02GB |
| notux-8x7b-v1.IQ3_S.gguf | IQ3_S | 19.03GB |
| notux-8x7b-v1.Q3_K_S.gguf | Q3_K_S | 19.03GB |
| notux-8x7b-v1.IQ3_M.gguf | IQ3_M | 19.96GB |
| notux-8x7b-v1.Q3_K.gguf | Q3_K | 21.0GB |
| notux-8x7b-v1.Q3_K_M.gguf | Q3_K_M | 21.0GB |
| notux-8x7b-v1.Q3_K_L.gguf | Q3_K_L | 22.51GB |
| notux-8x7b-v1.IQ4_XS.gguf | IQ4_XS | 23.63GB |
| notux-8x7b-v1.Q4_0.gguf | Q4_0 | 24.63GB |
| notux-8x7b-v1.IQ4_NL.gguf | IQ4_NL | 24.91GB |
| notux-8x7b-v1.Q4_K_S.gguf | Q4_K_S | 24.91GB |
| notux-8x7b-v1.Q4_K.gguf | Q4_K | 26.49GB |
| notux-8x7b-v1.Q4_K_M.gguf | Q4_K_M | 26.49GB |
| notux-8x7b-v1.Q4_1.gguf | Q4_1 | 27.32GB |
| notux-8x7b-v1.Q5_0.gguf | Q5_0 | 30.02GB |
| notux-8x7b-v1.Q5_K_S.gguf | Q5_K_S | 30.02GB |
| notux-8x7b-v1.Q5_K.gguf | Q5_K | 30.95GB |
| notux-8x7b-v1.Q5_K_M.gguf | Q5_K_M | 30.95GB |
| notux-8x7b-v1.Q5_1.gguf | Q5_1 | 32.71GB |
| notux-8x7b-v1.Q6_K.gguf | Q6_K | 35.74GB |
| notux-8x7b-v1.Q8_0.gguf | Q8_0 | 46.22GB |
language:
This model is a preference-tuned version of mistralai/Mixtral-8x7B-Instruct-v0.1 on the argilla/ultrafeedback-binarized-preferences-cleaned dataset using DPO (Direct Preference Optimization).
As of Dec 26th 2023, it outperforms Mixtral-8x7B-Instruct-v0.1 and is the top ranked MoE (Mixture of Experts) model on the Hugging Face Open LLM Leaderboard.
This is part of the Notus family of models and experiments, where the Argilla team investigates data-first and preference tuning methods like dDPO (distilled DPO). This model is the result of our first experiment at tuning a MoE model that has already been fine-tuned with DPO (i.e., Mixtral-8x7B-Instruct-v0.1).
We used a VM with 8 x H100 80GB hosted in runpod.io for 1 epoch (~10hr).
We used a new iteration of the Argilla UltraFeedback preferences dataset named argilla/ultrafeedback-binarized-preferences-cleaned.
The following hyperparameters were used during training:
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.4384 | 0.22 | 200 | 0.4556 | -0.3275 | -1.9448 | 0.7937 | 1.6174 | -405.7994 | -397.8617 | -1.3157 | -1.4511 |
| 0.4064 | 0.43 | 400 | 0.4286 | -0.2163 | -2.2090 | 0.8254 | 1.9927 | -408.4409 | -396.7496 | -0.7660 | -0.6539 |
| 0.3952 | 0.65 | 600 | 0.4275 | -0.1311 | -2.1603 | 0.8016 | 2.0291 | -407.9537 | -395.8982 | -0.6783 | -0.7206 |
| 0.3909 | 0.87 | 800 | 0.4167 | -0.2273 | -2.3146 | 0.8135 | 2.0872 | -409.4968 | -396.8602 | -0.8458 | -0.7738 |
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 73.18 |
| AI2 Reasoning Challenge (25-Shot) | 70.99 |
| HellaSwag (10-Shot) | 87.73 |
| MMLU (5-Shot) | 71.33 |
| TruthfulQA (0-shot) | 65.79 |
| Winogrande (5-shot) | 81.61 |
| GSM8k (5-shot) | 61.64 |