metadata
datasets:
- garage-bAInd/Open-Platypus
Instruction tune of Mistral-7B-v0.1 with Open-Platypus (fp16)
Overview
This is mistralai/Mistral-7B-v0.1, with instruction tuning performed with the garage-bAInd/Open-Platypus dataset.
This is a (merged) QLoRA fine-tune (rank 64).
The finetune was performed with 1x RTX 6000 Ada (~9 hours).
How to Use
As of writing, the Mistral architecture requires installation of transformers from source. With this done, load like any other model.
Benchmarks
ARC (25 shot): 62.80
Hellaswag (10 shot): 84.12
MMLU (5 shot): 64.20
Context Length - Relative Performance (wikitext perplexity)
| Context (tokens) | bhenrym14/mistral-7b-platypus-fp16 | bhenrym14/airoboros-l2-13b-2.1-YaRN-64k | bhenrym14/airophin-13b-pntk-16k-fp16 | bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16 | jondurbin/airoboros-l2-13b-gpt4-1.4.1 |
|---|---|---|---|---|---|
| 512 | 7.22 | 7.64 | 7.62 | 7.90 | 7.23 |
| 1024 | 6.04 | 6.15 | 6.20 | 6.17 | 5.85 |
| 2048 | 5.50 | 5.29 | 5.38 | 5.23 | 5.07 |
| 4096 | 5.05 | 4.93 | 5.08 | 4.91 | 4.77 |
| 8192 | 4.96 | 4.69 | 4.90 | Not Tested | 57.1 |
| 12000 | Not Tested | 4.53 | 4.82 | Not Tested | Not Tested |
- While the mistral model is very impressive for its size, particularly on benchmarks, the sliding window attention and/or model size impacts its competitiveness with other context extension techniques applied to larger llama2 and llama variants. Is this is more to do with sliding window attention or model size?
Prompting:
Model was trained with legacy airoboros <2.0 system prompt. See bhenrym14/airoboros-33b-gpt4-1.4.1-lxctx-PI-16384-fp16 model card for details.
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 56.89 |
| ARC (25-shot) | 63.05 |
| HellaSwag (10-shot) | 84.15 |
| MMLU (5-shot) | 64.11 |
| TruthfulQA (0-shot) | 45.07 |
| Winogrande (5-shot) | 78.53 |
| GSM8K (5-shot) | 17.36 |
| DROP (3-shot) | 45.92 |