mistral-7b-poetry

Mistral 7B — Poetry

A QLoRA fine-tune of Mistral 7B Instruct v0.3, trained on checkai/instruction-poems for instruction-driven poem generation — both explicit prompts ("write a poem about X") and open-ended continuations ("once upon a time a girl...").

Training details

Evaluation

Tested informally against the base model across six varied prompts — explicit topics (ocean, autumn/haiku, debugging code), open-ended story continuations, and an emotional introspective prompt — plus a general-capability sanity check.

Strengths. On narrative and descriptive prompts, the fine-tuned model consistently produced more concrete, original writing than the base model, which tended toward generic imagery and inspirational clichés. A prompt about debugging code, for example, produced a specific, genuinely funny narrative (staring at broken code, walking away, coming back to find a misspelled variable name) rather than the base model's abstract "coder's journey" framing. General reasoning capability outside poetry (e.g. basic math) remained intact — the model correctly switches to plain prose for non-creative questions rather than forcing poetic formatting everywhere.

Known quirks. Most responses open with a stray, non-sequitur preamble along the lines of "Here's a poem about the current date: [today's date]" before the actual poem begins — cosmetic, harmless, but noticeable. On at least one emotionally introspective, self-referential prompt, the model echoed the input back rather than generating new content; five other varied prompts did not reproduce this, so it appears to be an isolated weak spot rather than a general failure mode, but it's worth knowing about.

One theme worth flagging specifically: on very common poetic subjects (the ocean, in testing), output has occasionally closely echoed the phrasing of existing, well-known work rather than generating fully original text. Review output on classic/well-worn themes before reusing it, particularly for anything published or commercial.

This isn't a formal benchmark — it's a small, manually-run test set. Feedback from anyone who tries it more rigorously is genuinely welcome via this repo's discussions.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("agentbyumer/mistral-7b-poetry", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("agentbyumer/mistral-7b-poetry")

prompt = "Write a poem about the ocean."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=300)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

A GGUF-quantized version for local/CPU inference is available at agentbyumer/mistral-7b-poetry-GGUF.

Part of an ongoing series

One of several small, task-focused fine-tunes exploring what narrow specialization can improve in open models — covering function-calling, math reasoning, code generation, structured output, and creative writing so far.