qwen35-9b-takehome-expert-swe

LoRA adapter on Qwen/Qwen3.5-9B (revision c202236235762e1c871ad0ccb60c8ee5ba337b9a), part of a multi-expert training and model-merging experiment. Target: SWE-bench Pro v2-hard. Code, configs, logs and the full protocol: https://github.com/ktwu01/qwen36-35b-takehome

Training

QLoRA (nf4 base, bf16 compute), r=32, alpha=64, dropout 0, targets: full-attention q/k/v/o_proj, Gated DeltaNet in_proj_qkv/in_proj_z/out_proj, MLP gate/up/down_proj (language model only), lr 1e-4 cosine, 3% warmup, global batch 16 (union: 15), 16,384-token context, 1 epoch, assistant-only loss, seed 20261018.

Data

  • Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k (MIT)

Decontaminated with 13-gram overlap against all four benchmarks; see data/README.md in the code repo.

Use in the merge

This adapter's dense LoRA delta is one of the three task vectors of ktwu01/qwen35-9b-takehome-merged, merged with TIES (recipe below, configs/merge_final.json in the code repo; implementation src/takehome/merge.py).

{
 "name": "ties-d0.5-l1.0",
 "merge_args": "--method ties --experts outputs/swe outputs/term outputs/sci --density 0.5 --lam 1.0",
 "dev_mean_token_loss": 0.32781994345410975,
 "dev_token_loss": {
  "swe": 0.22026807381601063,
  "term": 0.1873204659077618,
  "sci": 0.5758712906385569
 },
 "rule": "lowest mean dev token loss over swe/term/sci among 3-expert merges; tie -> lower worst domain",
 "ranking": [
  {
   "name": "ties-d0.5-l1.0",
   "mean": 0.32782,
   "worst": 0.57587
  },
  {
   "name": "ties-d0.2-l1.0",
   "mean": 0.32953,
   "worst": 0.57779
  },
  {
   "name": "ties-d0.5-l0.7",
   "mean": 0.3317,
   "worst": 0.57579
  },
  {
   "name": "ta-l0.5",
   "mean": 0.33197,
   "worst": 0.58267
  },
  {
   "name": "dare-d0.5-l1.0",
   "mean": 0.3363,
   "worst": 0.58989
  },
  {
   "name": "ta-l0.7",
   "mean": 0.33852,
   "worst": 0.59366
  },
  {
   "name": "ta-l1.0",
   "mean": 0.36116,
   "worst": 0.62241
  }
 ]
}

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "ktwu01/qwen35-9b-takehome-expert-swe")   # pin revision=... from the candidate repo README
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B", revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a")

Evaluation

Status: final data (runners stopped 2026-10-19 20:20 CDT); Terminal-Bench seeds 2-3 incomplete, one complete seed per model. Experiment-specific protocol (apptainer containers, FP8 weights and KV cache, thinking on, temperature 1.0); not official leaderboard numbers. Full tables, per-task results and failure analysis: https://github.com/ktwu01/qwen36-35b-takehome (README section 6, analysis/failures.md).

Benchmark This model Base (Qwen3.5-9B)
SWE-bench Pro v2-hard (HARD-51, resolved %) 0.00 5.88

Experts were evaluated on their target benchmarks only.

9 of 51 SWE-Pro trials end in RepeatedFormatError: replies without a closing think tag are returned by vLLM's qwen3 reasoning parser as reasoning_content only.

License and limitations

Apache-2.0 (as the base model). Training data licenses above; OpenScienceReasoning-2 is CC-BY-4.0 and requires attribution to NVIDIA. Research artifact: evaluated under the experiment-specific protocol in the code repository (apptainer environment, FP8 serving); scores are not official leaderboard results.

Downloads last month
15
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ktwu01/qwen35-9b-takehome-expert-swe

Finetuned
Qwen/Qwen3.5-9B
Adapter
(1,121)
this model