Llama-Locali-adapters

The other adapters from the Locali fine-tuning experiments. They are here so the comparisons in the project report can be reproduced. The adapter I actually use is Llama-Locali-Qwen2.5-Coder-1.5B-QLoRA. The name prefix is explained on that page (Llama 3.1 licence).

Folder	Base model	Method	Data	Training time
qwen2.5-coder-1.5b-qlora-v1	Qwen2.5-Coder-1.5B-Instruct 2e1fd39	QLoRA (NF4)	mixture v1	1 h 58 min
qwen2.5-coder-1.5b-lora16-v2	Qwen2.5-Coder-1.5B-Instruct 2e1fd39	LoRA on the 16-bit base	mixture v2	2 h 17 min
qwen2.5-coder-1.5b-rslora-v2	Qwen2.5-Coder-1.5B-Instruct 2e1fd39	QLoRA + rank-stabilised scaling	mixture v2	2 h 34 min
phi-4-mini-qlora-v2	unsloth/Phi-4-mini-instruct 75becc4 (same weights as microsoft/Phi-4-mini-instruct)	QLoRA (NF4)	mixture v2	5 h 58 min

All of them use r = 16, alpha = 16, all projections, 2 epochs, lr 2e-4, seed 3407, one Kaggle T4. Mixture v1 is a subset of locali-mix-v2: the rows without the T4 traces, the second family of bug operators and the extra clean reviews.

What I found, in short:

• v1 → v2: the v1 adapter lost 4.6 points on CRUXEval-O. Adding execution-trace data in v2 removed that loss. v1 also only improved on bug types close to its training faults.
• QLoRA vs 16-bit LoRA vs rsLoRA (same data): no consistent winner. Quantizing the frozen base to NF4 cost nothing measurable and saved 36 % of peak memory. Each variant beat the others on one or two tasks, by amounts that one seed cannot separate from noise.
• Phi-4-mini: it gained where the training answers were better than what it already did (in-distribution repair +30 points, review +22, secure code generation +17). It lost where base Phi was already better than our targets (repair in our prompt format −22, in-distribution test writing −11). It also became worse at using the function name a prompt asks for, which costs it 6 points on MBPP+.

Tables and significance tests: experiments/results/finetune_v2/README.md and experiments/results/finetune_qwen25_qlora_v1/README.md in the project repository.

Load one with subfolder:

from peft import PeftModel
model = PeftModel.from_pretrained(base_model, "MedGm/Llama-Locali-adapters", subfolder="qwen2.5-coder-1.5b-rslora-v2")

Built with Llama. The Qwen adapters are Apache-2.0. The Phi-4-mini adapter follows its base model's MIT licence. The Llama 3.1 Community License naming and attribution terms apply to all of them because of the training data.