Fine-Tuning a 7B Advisor on Free-Tier GPUs: An Adapter-Handoff Recipe and a Synthetic-Data Reliability Caution

A team fine-tuned a 7-billion-parameter language model on free-tier GPUs using a three-epoch QLoRA method, but the resulting model performed worse on advising quality than the unmodified base model, according to a preprint posted to arXiv [1]. The researchers detail a practical recipe for completing a multi-epoch fine-tune of Mistral-7B-Instruct-v0.3 across two free-tier 16 GB GPUs — a Tesla P100 and then a T4 — by checkpointing only the small 41.9-million-parameter LoRA adapter and resuming on the second machine [1]. The adapter-only handoff, which does not require transferring optimizer or scheduler state, makes the binding constraint per-step VRAM and per-session wall-clock rather than aggregate compute [1]. The work targets resource-constrained settings such as Kaggle and Colab, where multi-epoch runs routinely exceed session limits [1]. A blind evaluation against the un-fine-tuned base model returned a cautionary result. The fine-tuned model scored higher on similarity to the synthetic training distribution, with a BERTScore F1 increase of 0.063, but lower on advising quality [1]. A blind LLM-as-judge preferred the base model on 46% of prompts versus 18% for the fine-tuned model [1]. A source-verified factuality audit found four confident errors from the fine-tuned model on policy-sensitive topics, while the base model made zero [1]. Auditing the training data with the same method showed the errors were not a fine-tuning artifact. Each audited error was already present in the Gemini-generated training answers, and a random-sample audit found verifiable errors in 28-40% of responses (single-judge, n=40) [1]. The authors attribute the errors to the synthetic-data pipeline rather than the adapter-handoff method [1]. The dataset, adapter, cross-GPU notebooks, and full evaluation harness have been released so that every result can be reproduced on a single 16 GB GPU [1]. The paper appears on arXiv as version four of a preprint first submitted in April 2025 and most recently revised in June 2026 [1]. The Hugging Face Hub, which indexes arXiv papers and links them to associated models, datasets, and demos, provides a mechanism for the community to discuss such work and for authors to claim verified authorship [4]. The platform also supports embedding interactive demos directly alongside papers on arXiv abstract pages, allowing users to try implementations without writing code [5].