minihf_evaluator_openllama_7b

minihf_evaluator_openllama_7b is a LoRA instruct fine-tune of OpenLLaMA 7B.

The sequence <|end|> was used to separate the prompt and response. The correct way to prompt the model is: Does 2 + 2 = 4?<|end|>. The tokenizer will prepend a BOS token (<s>) by default. The response will end with an EOS token (</s>).

Training procedure

minihf_evaluator_openllama_7b was fine-tuned for 100,000 examples on 90% Muennighoff/flan / 10% databricks/databricks-dolly-15k using batch size 4 per GPU on 8 40GB A100 GPUs. Examples where the prompt and response would not fit into 2,048 tokens were dropped. The fine-tuning was done using the following command:

accelerate launch make_evaluator.py --output-dir minihf_evaluator_openllama_7b

The following bitsandbytes quantization config was used during training:

• 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 0.4.0.dev0