ScalingIntelligence/qwen3.6-trace-semantic-logic-precision-v1

Qwen3.6-27B TRACE LoRA — Target Capability: semantic-logic-precision

GRPO-trained LoRA adapters on top of Qwen/Qwen3.6-27B, produced by the TRACE (Test-time Reinforcement Learning Adaptation via Capability Environments) pipeline.

Target capability

semantic-logic-precision was identified via open-ended capability mining (phase 2 of TRACE): Qwen3.6 itself read the 123 failed SWE-bench Verified trajectories from its own baseline run and proposed a 5-category taxonomy of its weaknesses. semantic-logic-precision was the largest category at 61% of failures:

Implements a logically incorrect algorithm — misunderstanding domain protocols, mathematical rules, or language semantics. Code is syntactically valid but solves the wrong problem.

Training data

Synthetic single-turn SEARCH/REPLACE scenarios (phase 3 v4 + v5 mutation):

Scenario	Base success rate	Description
pid-autotuner-ku-formula-fix	0.30	Wrong Ziegler-Nichols Ku formula (extra * 2 in denominator)
pid-autotuner-ku-formula-fix_mut1_1	0.50	Same bug, surface code disguised: identifiers renamed, decoy edit sites, misleading "logic is correct" comment

Training setup

• Trainer: GRPO with rank-16 LoRA (q_proj, k_proj, v_proj, o_proj)
• Group size: 8 rollouts per group (DS-retry on constant-reward groups)
• Inference backend: vLLM TP=4 with hot-reload LoRA after each policy step
• Hardware: 4×H100 80GB
• Total iterations: 130 (saturated around iter ~30-50)
• Per-iter wall time: ~70-100 sec

Checkpoints

Subfolder	When taken	Notes
iter_30/	Mid-learning	Generalizable signal, low overfit risk
iter_50/	Recommended	Best balance of learning vs overfit
iter_100/	Near-saturation	High training-set performance, possibly overfit
iter_130/	Final	Peak training reward, may not transfer better than iter_50

Loading

vLLM (production / eval)

python3 -c "
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='ScalingIntelligence/qwen3.6-trace-semantic-logic-precision-v1',
    allow_patterns='iter_50/*',
    local_dir='/tmp/trace_lora',
)
"

curl -X POST http://localhost:8000/v1/load_lora_adapter \
  -H 'Content-Type: application/json' \
  -d '{"lora_name": "trace_iter50", "lora_path": "/tmp/trace_lora/iter_50"}'

PEFT + transformers

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.6-27B", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(
    base,
    "ScalingIntelligence/qwen3.6-trace-semantic-logic-precision-v1",
    subfolder="iter_50",
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B")

Merged model (no PEFT at inference)

m = PeftModel.from_pretrained(base, REPO, subfolder="iter_50")
merged = m.merge_and_unload()
merged.save_pretrained("./qwen3.6-trace-slp-iter50-merged")

Evaluation

SWE-bench Verified Pass@1, mini-swe-agent + Modal sandboxes, vLLM @ 128k:

Model	Pass@1	Notes
Qwen3.6-27B baseline	58.2%	phase 1, untrained
+ iter_50 LoRA	pending	phase 6 in progress
+ iter_100 LoRA	pending	will run after iter_50 finishes

Caveats

• Training data is 2 single-turn scenarios — a narrow slice of the target capability. Transfer to SWE-bench's diverse multi-file, multi-turn bugs is the open research question this artifact tests.
• Without a KL penalty (--kl-coef 0.0 in training), late checkpoints (iter_100, iter_130) drift further from the base policy than typically desirable. iter_30 / iter_50 may transfer better in practice.
• Mutations are stronger than the base scenarios (forced search-misdirection via renamed identifiers, decoys, and a contradictory comment); however even the mutated scenarios are far simpler than real open-source bugs.

Citation / acknowledgements

Produced as part of TRACE pipeline development at Stanford ScalingIntelligence Lab.