Downloads · 30 days
12
23% of all-time downloads
ahmedikram/SWE-Qwen-qwen3-14b-higher_rank_14b
SWE-Qwen-qwen3-14b-higher_rank_14b is a text generation model from ahmedikram. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
SWE-Qwen-qwen3-14b-higherrank14b is a QLoRA fine-tune of Qwen3-14B for automated software-issue resolution: given a GitHub issue, it generates a patch that flips FAILTOPASS tests and preserves PASSTOPASS tests inside…
Downloads · 30 days
12
23% of all-time downloads
All-time downloads
52
Public
Repo size
525 MB
Likes
0
Public
Click a slice to open those files.
.safetensors514 MB · 98%
From the Hugging Face model README
higher_rank_14b (software-issue resolution)SWE-Qwen-qwen3-14b-higher_rank_14b is a QLoRA fine-tune of Qwen3-14B for automated software-issue resolution: given a GitHub issue, it generates a patch that flips FAIL_TO_PASS tests and preserves PASS_TO_PASS tests inside the real repository.
Qwen/Qwen3-14B (Apache-2.0)higher_rank_14b — LoRA r=32, α=64, lr 2e-5, 1 epoch, bf16, paged_adamw_8bit, cosine (warmup 0.03), weight decay 0.01, max seq len 4,096, packing onExecution-based evaluation inside official SWE-bench Docker images on a 100-instance sample of the 2,313-instance golden set (statistics: Wilson 95% CI, McNemar + paired bootstrap for champion selection):

| Variant | F2P (95% Wilson CI) | P2P | Latency | Flaky | Note |
|---|---|---|---|---|---|
| Qwen3-14B (base) | 2.46% (0.8–7.7%) | 28.54% | 10.04 s | 0.06% | rejected: p2p<90%, f2p<15% |
| baseline_14b (r16) | 12.30% (7.2–20.2%) | 85.90% | 9.47 s | 0.03% | rejected: p2p<90% |
| higher_rank_14b | 17.20% (11.1–25.8%) | 90.10% | 8.92 s | 0.01% | champion |
| higher_lr_14b | 14.80% (9.1–23.1%) | 87.60% | 9.12 s | 0.02% | rejected: p2p<90% |
Champion vs base: F2P 7.0× (2.46% → 17.20%), P2P +61.6pt (28.54% → 90.10%); McNemar p ≈ 6e-05, paired-bootstrap 95% CI lower bound > 0.
baseline_14b, higher_rank_14b, higher_lr_14b) on one A100-80GB budget; ~4,214 s (~1.2 h) per variant, champion higher_rank_14b final train loss 0.5843model-qwen3-14b-{variant})Serve with vLLM (AWQ base + per-request LoRA swap; OpenAI-compatible):
from vllm import LLM, SamplingParams
llm = LLM(model="Qwen/Qwen3-14B-AWQ", enable_lora=True)
params = SamplingParams(temperature=0.2, max_tokens=2048)
out = llm.generate(["<full prompt: system + issue>"], params, lora_request=...) # adapter via LoRARequest
Fine-tune shares the base model's license (Apache-2.0). Benchmarks target the curated SWE-bench subset (Python); latency measured on warm Modal A100-80GB pull-through caching. Eval harness and full audit trail: see the SWE-Qwen repository.