Try generating higher volume of test data (via Mistral 7B ?)

[mrseanryan]

Approach:

• use a larger context window

• repeat and concatenate, with higher temperature

• multiply programmatically (post-process)

---

LLM - Mistral 7B

Context size: a sliding 4K window - see https://huggingface.co/mistralai/Mistral-7B-v0.1/discussions/4

There are many flavours of Mistral-7B:

1. regular mistral-7B, quantized

• TheBloke/Mistral-7B-v0.1-GGUF = https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF

2. cognitivecomputations/dolphin-2.6-mistral-7b-dpo

• good for coding, has 16K context (note: is not 'ethically trained'
• https://huggingface.co/cognitivecomputations/dolphin-2.6-mistral-7b-dpo
• quantized = https://huggingface.co/TheBloke/dolphin-2.6-mistral-7B-dpo-GPTQ

3. • mistral-7B with larger context window

• TheBloke_Writing_Partner_Mistral_7B-GPTQ_gptq-8bit-32g-actorder_True = https://huggingface.co/TheBloke/Writing_Partner_Mistral_7B-GPTQ/tree/gptq-8bit-32g-actorder_True

4 - instruct prompting, and newer:

• https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GPTQ

5. a base version, suitable for FT

• https://huggingface.co/TheBloke/mistral-ft-optimized-1227-GPTQ

6. hermes-2.5 - mistral-7B

• https://huggingface.co/mlabonne/NeuralHermes-2.5-Mistral-7B
• quantized = https://huggingface.co/TheBloke/NeuralHermes-2.5-Mistral-7B-AWQ
  • or https://huggingface.co/NousResearch/Nous-Hermes-2-Mistral-7B-DPO-GGUF

7. phi from Microsoft

• phi-1.5 (from Microsoft) - Open LLM leaderboard
  = https://huggingface.co/microsoft/phi-1_5
• phi-2 = https://huggingface.co/microsoft/phi-2