plan-then-code-lora-14b

LoRA adapter for Qwen/Qwen2.5-Coder-14B, trained to "plan before coding" — first emit an I/O contract + 3–8 steps of pseudocode, then the code.

On 324 held-out algorithm problems, pass@1 improves from 36.7% to 45.7% (+9.0pp, McNemar p<0.001), compile rate from 99.4% to 100%.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen2.5-Coder-14B"
adapter = "byesher/plan-then-code-lora-14b"   # or a local adapter path

tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)

messages = [{"role": "user", "content": "【题目】\n{题目}\n\n请先写出分步伪代码（算法计划），再写出完整可运行的 Python 程序。"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Results

Three metrics (324 held-out algorithm problems)

Setting	pass@1	compile	case_pass
Bare Coder, chat prompt	6.2% (20/324)	26.2% (85/324)	6.6% (256/3894)
Bare Coder, comment-completion	36.7% (119/324)	99.4% (322/324)	34.1% (1327/3894)
SFT (plan-then-code)	45.7% (148/324)	100% (324/324)	36.7% (1429/3894)

Relative to the "comment-completion" baseline, SFT improves pass@1 by +9.0pp (McNemar test, p < 0.001).

Failure breakdown (324 problems, by failure reason)

Setting	Passed	Compile error	Wrong output (mismatch)	Runtime error	Timeout
Bare Coder, chat prompt	20	239	41	22	2
Bare Coder, comment-completion	119	2	172	26	5
SFT (plan-then-code)	148	0	167	8	1

Training

• Base model: Qwen2.5-Coder-14B (base)
• LoRA: r=8, alpha=16, lr=1e-4, 1 epoch, bf16
• Training data: 2,372 samples (closed-loop verification, 81.4% acceptance)
• Generation: single-turn, one call per problem (plan + code together)

Repository

• Code & full documentation: byesher/plan-then-code

Framework versions

• PEFT 0.21.2