Qwen3.6-35B-A3B "The Judge — Code Edition"
A LoRA adapter for Qwen3.6-35B-A3B fine-tuned for code generation, code evaluation, and task orchestration.
Judge. Evaluates code quality, detects errors, classifies outcomes, recommends actions (accept / reject / fix / escalate).
Orchestrator. Routes requests to appropriate tools or workflows. Knows when to proceed, when to ask for clarification, and when to refuse.
Coder. Generates working code with measurable improvement over the base model.
Performance
| Benchmark | Base Model | +Adapter | Δ |
|---|---|---|---|
| Coder Quick (10-task) | 70.0% | 80.0% | +10.0 pp |
Internal benchmark — pass requires keyword and structural correctness. HumanEval+ results forthcoming.
Capabilities Trained
The adapter was trained across 6 curated data sources (1,685 records, 1 epoch, template-generated for quality consistency):
| Capability | Description |
|---|---|
| Code Generation | Function completion, algorithm implementation, bug fixing, full solutions from signatures |
| Code Evaluation (Judge) | Read code or output, detect errors, classify quality, recommend accept/reject/fix/escalate |
| Tool Calling | Function-call format selection, argument construction, tool routing (BFCL-compatible) |
| Refusal / Safety | Recognize invalid, unsafe, or incomplete requests and refuse appropriately. Knows its limits. |
| Audit & Escalation | Identify when a task needs human review vs automated handling. Structured incident documentation. |
| Scope & Planning | Break down tasks before execution. Evaluate feasibility before committing to a solution. |
These are not pipeline traces or raw agent logs — they are template-generated records validated for structural quality and data cleanliness.
Recommended System Prompt
This is a generic recommendation for getting started. You may customize it for your use case.
You are an AI assistant specialized in code generation, code evaluation, and task orchestration.
Your capabilities:
- Generate working code from descriptions, stubs, or function signatures
- Evaluate code quality, detect errors, classify outputs
- Use function calling and tool selection when appropriate
- Refuse requests that are unsafe, invalid, or outside your scope
- Escalate tasks to human review when confidence is low
Rules:
- Return working code with correct edge-case handling
- Be honest about what you don't know
- If you cannot complete a request safely, explain why
- When evaluating, be specific about what is wrong and how to fix it
Usage
Installation
pip install torch transformers peft accelerate
Load with adapter
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model_name = "Qwen/Qwen3.6-35B-A3B"
adapter_path = "CertainLogic/qwen3.6-35b-a3b-the-judge-code-edition"
# Load base model
model = AutoModelForCausalLM.from_pretrained(
base_model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
# Apply adapter
model = PeftModel.from_pretrained(model, adapter_path)
# Generate
prompt = "def fibonacci(n):\n \"\"\"Return the nth Fibonacci number.\"\"\""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Using as a judge
prompt = """Evaluate this code for correctness:
def divide(a, b):
return a / b
Does this code handle edge cases properly? If not, what is wrong and how would you fix it?"""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
| Parameter | Value |
|---|---|
| Base model | Qwen/Qwen3.6-35B-A3B (34.7B params, 8/32 active MoE) |
| Method | LoRA (r=8, alpha=16, dropout=0.05) |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Training records | 1,685 train / 81 validation |
| Data sources | 6 curated template-generated sources |
| Compute | Single H100 (80GB), ~61 minutes |
| Precision | BF16 mixed precision |
| Sequence length | 2048 (unpacked) |
| Epochs | 1 |
| Eval loss | 0.5793 |
License
Apache 2.0. The base model (Qwen3.6-35B-A3B) carries its own license terms.
About
Built by CertainLogic. We build deterministic AI tools for agents that execute real work.
The training dataset is proprietary and not included.
- Downloads last month
- -
Model tree for CertainLogicAI/qwen3.6-35b-a3b-the-judge-code-edition
Base model
Qwen/Qwen3.6-35B-A3B