A LoRA adapter fine-tuned on the CodeK v1 dataset: a reasoning-first, pedagogical coding dataset. Teaches decomposition, bug diagnosis, contrast reasoning, and hypothesis-driven thinking about code.

v0 Eval Results (Pass 2 ground-truth, 50 seeds)

Model	Pass@1
Base (Qwen2.5-Coder-7B-Instruct)	64%
LoRA checkpoint-800	58%

6% regression on bug diagnosis. LoRA wins on 2/50 seeds (more direct, correct), base wins on 5/50 (LoRA misidentifies function or pattern-matches to training data). See BASELINE_V0.md in the dataset repo for full analysis.

Training

Setting	Value
Base model	`Qwen/Qwen2.5-Coder-7B-Instruct`
Method	LoRA (RS-LoRA)
Rank / Alpha	16 / 32
Dropout	0.05
Epochs	3
Batch (effective)	8
Learning rate	2e-4
Train pairs	2,351
Best eval loss	0.0583 (step 528)
Checkpoint used	checkpoint-800 (eval loss 0.061)
Hardware	RunPod A100 80GB, 59 min

Dataset

mechramc/codek-v1 -- 201 seeds, 4 augmentation passes, 2,613 ShareGPT pairs.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")
model = PeftModel.from_pretrained(base, "mechramc/codek-qwen2.5-coder-7b-lora")
tokenizer = AutoTokenizer.from_pretrained("mechramc/codek-qwen2.5-coder-7b-lora")