✅ Creative assistants
✅ Educational explanations in poetic form
✅ Fun chatbots
✅ Style-controlled generation experiments
❌ Not intended for strict factual QA without creative framing

How to Load the Model

Developed by: khemn
License: Apache 2.0
Base model: unsloth/Phi-3-mini-4k-instruct-bnb-4bit
Fine-tuning method: LoRA (via Unsloth)
Training framework: TRL + Unsloth
Precision: 4-bit (bnb)

Using Unsloth (recommended)

python

from unsloth import FastLanguageModel

model_name = "khemn/poetic-assistant-phi3"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=2048,
    load_in_4bit=True
)

FastLanguageModel.for_inference(model)

Parameter	Recommended Value	Reason
`temperature`	0.7–0.9	Creative but coherent
`do_sample`	True	Enables poetic variation
`top_p`	0.9	Prevents nonsense
`max_new_tokens`	128–256	Enough space for poems

Training Details (High Level)

Dataset format: Instruction → Poetic response
LoRA fine-tuning with Unsloth
4-bit quantization for efficiency
Optimized for fast iteration and low VRAM usage