Examples

As you can see from examples bellow the output is far from ideal, and far from simple GPT/LLama2 prompt without finetuning.

Quality issues:

• the corrected sentences are containing major errors
• the explanations of corrections are wrong The low quality is probably caused by very low volume of training data, and small LLM size.

Interesting observation: The LLM react 100% on the learned example 1 (exactly from training data)

Example 1

Input

''### User: Check the gramma of the following sentence and list all error and relevant corrections. Sentence:Mir es geht gut, danke#### AI:Correct version of the sentence:

Output

"Mir geht es gut, danke."

Repairs:

• "Mir es geht" the verb should be placed in the second position "Mir geht es" - because of the correct word order.
• The sentence lacks a period at the end.#### End of repair list for this sentence.#### End of topic, repair list and the sentence

Example 2

Input

''### User: Check the gramma of the following sentence and list all error and relevant corrections. Sentence:Er gehen dort, aber ich muss ander geneh.#### AI:Correct version of the sentence:

Output

Er macht dort dorte, aber ich muss einmal dorte.

Repairs:

• Instead of "Er gehen dort" it should be "Er macht dort" - the correct verb form is "macht"
• Instead of "ich muss" it should be "ich muss einmal" - the verb "einmal" is used in the future in German.
• The verb "dorte" must be conjugated in the third person singular "dorte".#### End of Repairs:
• Instead of "Er gehen dort" it should be "Er macht dort" - the correct verb form is "macht".

Training procedure

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: False
• bnb_4bit_compute_dtype: float16

Framework versions

• PEFT 0.4.0