'
- f'
'
f'{slogan}'
'
'
f'{get_text("disclaimer")}'
@@ -348,7 +345,7 @@ def main():
setup_seed(random_seed)
tools = [t for t in TOOLS if t['function']['name'] in st.session_state.get('selected_tools', [])] or None
- sys_prompt = [] if tools else [{"role": "system", "content": "你是MiniMind，一个乐于助人、知识渊博的AI助手。请用完整且友好的方式回答用户问题。"}]
+ sys_prompt = [] if tools else [{"role": "system", "content": "你是一个用于轻量语言模型实验的AI助手。请准确、完整且友好地回答用户问题，不要虚构开发者或个人身份。"}]
st.session_state.chat_messages = sys_prompt + st.session_state.chat_messages[-(st.session_state.history_chat_num + 1):]
template_kwargs = {"tokenize": False, "add_generation_prompt": True}
if st.session_state.get('enable_thinking', False):
diff --git a/tests/fixtures/code_predictions.jsonl b/tests/fixtures/code_predictions.jsonl
new file mode 100644
index 0000000..2b5c7e0
--- /dev/null
+++ b/tests/fixtures/code_predictions.jsonl
@@ -0,0 +1 @@
+{"task_id":"add-two","candidates":["def add(a, b):\n return a + b","def add(a, b):\n return a - b"]}
diff --git a/tests/fixtures/code_rl_predictions.jsonl b/tests/fixtures/code_rl_predictions.jsonl
new file mode 100644
index 0000000..61267d1
--- /dev/null
+++ b/tests/fixtures/code_rl_predictions.jsonl
@@ -0,0 +1,4 @@
+{"task_id":"reverse-string","candidates":["def reverse_string(text):\n return text[::-1]","def reverse_string(text):\n return text"]}
+{"task_id":"valid-parentheses","candidates":["def valid_parentheses(text):\n pairs = {')': '(', ']': '[', '}': '{'}\n stack = []\n for char in text:\n if char in '([{':\n stack.append(char)\n elif not stack or stack.pop() != pairs[char]:\n return False\n return not stack","def valid_parentheses(text):\n return True"]}
+{"task_id":"binary-search","candidates":["def binary_search(values, target):\n left, right = 0, len(values) - 1\n while left <= right:\n mid = (left + right) // 2\n if values[mid] == target:\n return mid\n if values[mid] < target:\n left = mid + 1\n else:\n right = mid - 1\n return -1","def binary_search(values, target):\n return -1"]}
+{"task_id":"max-subarray","candidates":["def max_subarray(values):\n best = current = values[0]\n for value in values[1:]:\n current = max(value, current + value)\n best = max(best, current)\n return best","def max_subarray(values):\n return sum(values)"]}
diff --git a/tests/fixtures/code_tasks.jsonl b/tests/fixtures/code_tasks.jsonl
new file mode 100644
index 0000000..e3c7461
--- /dev/null
+++ b/tests/fixtures/code_tasks.jsonl
@@ -0,0 +1 @@
+{"task_id":"add-two","prompt":"Implement add(a, b) and return the sum.","entry_point":"add","tests":[{"args":[1,2],"expected":3},{"args":[-5,5],"expected":0}]}
diff --git a/tests/test_code_agent.py b/tests/test_code_agent.py
new file mode 100644
index 0000000..5a815ce
--- /dev/null
+++ b/tests/test_code_agent.py
@@ -0,0 +1,47 @@
+from code_agent import CodeTask, ExecutionFeedbackAgent, TestCase as CodeTestCase
+
+
+TASK = CodeTask(
+ task_id="maximum",
+ prompt="Implement maximum(values), returning the largest integer.",
+ entry_point="maximum",
+ tests=(
+ CodeTestCase(args=([1, 9, 3],), expected=9),
+ CodeTestCase(args=([-5, -2, -8],), expected=-2),
+ ),
+)
+
+
+class ScriptedGenerator:
+ def __init__(self, responses):
+ self.responses = iter(responses)
+ self.prompts = []
+
+ def __call__(self, prompt):
+ self.prompts.append(prompt)
+ return next(self.responses)
+
+
+def test_agent_repairs_solution_from_hidden_test_feedback():
+ generator = ScriptedGenerator(
+ [
+ "```python\ndef maximum(values):\n return values[0]\n```",
+ "```python\ndef maximum(values):\n return max(values)\n```",
+ ]
+ )
+ result = ExecutionFeedbackAgent(generator, max_attempts=3).run(TASK)
+
+ assert result.success
+ assert len(result.attempts) == 2
+ assert result.attempts[0].verification.pass_rate == 0.0
+ assert "hidden test failed" in generator.prompts[1]
+ assert "expected 9" not in generator.prompts[1]
+ assert result.final.reward.total == 1.0
+
+
+def test_agent_stops_at_attempt_limit():
+ generator = ScriptedGenerator(["def maximum(values):\n return 0"] * 2)
+ result = ExecutionFeedbackAgent(generator, max_attempts=2).run(TASK)
+
+ assert not result.success
+ assert len(result.attempts) == 2
diff --git a/tests/test_code_backend.py b/tests/test_code_backend.py
new file mode 100644
index 0000000..3ccacfe
--- /dev/null
+++ b/tests/test_code_backend.py
@@ -0,0 +1,42 @@
+from types import SimpleNamespace
+
+import pytest
+
+from code_agent import OpenAICompatibleGenerator
+
+
+class FakeCompletions:
+ def __init__(self, content):
+ self.content = content
+ self.kwargs = None
+
+ def create(self, **kwargs):
+ self.kwargs = kwargs
+ message = SimpleNamespace(content=self.content)
+ return SimpleNamespace(choices=[SimpleNamespace(message=message)])
+
+
+def test_openai_backend_builds_non_streaming_request():
+ completions = FakeCompletions("```python\ndef solve():\n return 1\n```")
+ client = SimpleNamespace(chat=SimpleNamespace(completions=completions))
+ generator = OpenAICompatibleGenerator(
+ client=client,
+ model="local-minimind",
+ temperature=0.2,
+ max_tokens=128,
+ open_thinking=True,
+ )
+
+ assert generator("solve it").startswith("```python")
+ assert completions.kwargs["messages"][0]["content"] == "solve it"
+ assert completions.kwargs["stream"] is False
+ assert completions.kwargs["extra_body"] == {"open_thinking": True}
+
+
+def test_openai_backend_rejects_empty_completion():
+ completions = FakeCompletions("")
+ client = SimpleNamespace(chat=SimpleNamespace(completions=completions))
+ generator = OpenAICompatibleGenerator(client=client)
+
+ with pytest.raises(RuntimeError, match="empty completion"):
+ generator("solve it")
diff --git a/tests/test_code_evaluation.py b/tests/test_code_evaluation.py
new file mode 100644
index 0000000..9901ee4
--- /dev/null
+++ b/tests/test_code_evaluation.py
@@ -0,0 +1,54 @@
+import pytest
+
+from code_agent import (
+ CodeTask,
+ TestCase as CodeTestCase,
+ VerificationStatus,
+ aggregate_reports,
+ estimate_pass_at_k,
+ evaluate_candidates,
+)
+
+
+TASK = CodeTask(
+ task_id="square",
+ prompt="Implement square(x).",
+ entry_point="square",
+ tests=(
+ CodeTestCase(args=(2,), expected=4),
+ CodeTestCase(args=(-3,), expected=9),
+ ),
+)
+
+
+def test_pass_at_k_estimator():
+ assert estimate_pass_at_k(10, 0, 1) == 0.0
+ assert estimate_pass_at_k(10, 1, 10) == 1.0
+ assert estimate_pass_at_k(10, 2, 1) == pytest.approx(0.2)
+ with pytest.raises(ValueError):
+ estimate_pass_at_k(2, 1, 3)
+
+
+def test_evaluation_reports_execution_metrics():
+ candidates = [
+ "def square(x):\n return x * x",
+ "def square(x):\n return x + x",
+ "def square(x)\n return x * x",
+ ]
+ report = evaluate_candidates(TASK, candidates, ks=(1, 3))
+
+ assert report.correct_count == 1
+ assert report.compile_rate == pytest.approx(2 / 3)
+ assert report.execution_rate == pytest.approx(2 / 3)
+ assert report.test_pass_rate == pytest.approx(0.5)
+ assert report.pass_at_k["pass@1"] == pytest.approx(1 / 3)
+ assert report.pass_at_k["pass@3"] == 1.0
+ assert report.results[2].status == VerificationStatus.SYNTAX_ERROR
+
+ summary = aggregate_reports([report])
+ assert summary["candidate_count"] == 3
+ assert summary["status_counts"] == {
+ "passed": 1,
+ "syntax_error": 1,
+ "wrong_answer": 1,
+ }
diff --git a/tests/test_code_reward_batch.py b/tests/test_code_reward_batch.py
new file mode 100644
index 0000000..798e833
--- /dev/null
+++ b/tests/test_code_reward_batch.py
@@ -0,0 +1,25 @@
+import pytest
+
+from code_agent import CodeTask, TestCase as CodeTestCase, score_code_batch
+
+
+TASK = CodeTask(
+ task_id="identity",
+ prompt="Implement identity(x).",
+ entry_point="identity",
+ tests=(CodeTestCase(args=(7,), expected=7),),
+)
+
+
+def test_scores_task_major_grouped_responses():
+ rewards = score_code_batch(
+ [TASK],
+ ["def identity(x):\n return x", "def identity(x):\n return 0"],
+ num_generations=2,
+ )
+ assert [reward.total for reward in rewards] == [1.0, 0.0]
+
+
+def test_rejects_misaligned_group_size():
+ with pytest.raises(ValueError, match="expected 2 responses"):
+ score_code_batch([TASK], ["def identity(x): return x"], num_generations=2)
diff --git a/tests/test_code_rl_dataset.py b/tests/test_code_rl_dataset.py
new file mode 100644
index 0000000..9c19ab8
--- /dev/null
+++ b/tests/test_code_rl_dataset.py
@@ -0,0 +1,31 @@
+import json
+
+from code_agent import CodeRLDataset, collate_code_rl
+
+
+class FakeTokenizer:
+ def apply_chat_template(self, messages, **kwargs):
+ return messages[0]["content"] + f"\nthinking={kwargs['open_thinking']}"
+
+
+def test_code_rl_dataset_builds_prompt_and_preserves_tests(tmp_path):
+ path = tmp_path / "tasks.jsonl"
+ path.write_text(
+ json.dumps(
+ {
+ "task_id": "inc",
+ "prompt": "Implement inc(x).",
+ "entry_point": "inc",
+ "tests": [{"args": [1], "expected": 2}],
+ }
+ )
+ + "\n",
+ encoding="utf-8",
+ )
+ dataset = CodeRLDataset(path, FakeTokenizer(), thinking_ratio=0.0)
+ sample = dataset[0]
+
+ assert "Required entry point: inc" in sample["prompt"]
+ assert sample["task"].tests[0].expected == 2
+ batch = collate_code_rl([sample])
+ assert batch["tasks"][0].task_id == "inc"
diff --git a/tests/test_code_verifier.py b/tests/test_code_verifier.py
new file mode 100644
index 0000000..c7a7efa
--- /dev/null
+++ b/tests/test_code_verifier.py
@@ -0,0 +1,57 @@
+from code_agent import (
+ CodeTask,
+ TestCase as CodeTestCase,
+ VerificationStatus,
+ extract_python_code,
+ score_verification,
+ verify_code,
+)
+
+
+TASK = CodeTask(
+ task_id="add-two",
+ prompt="Implement add(a, b).",
+ entry_point="add",
+ tests=(
+ CodeTestCase(args=(1, 2), expected=3),
+ CodeTestCase(args=(5, 0), expected=5),
+ ),
+)
+
+
+def test_extracts_fenced_code_after_thinking():
+ response = "