topoexploit 0.2.0__tar.gz → 0.3.0__tar.gz

{topoexploit-0.2.0 → topoexploit-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: topoexploit
-Version: 0.2.0
+Version: 0.3.0
 Summary: TopoExploit: complex-valued spectral language model for security tasks (vulnerability analysis, patch analysis, exploit development), trained on ExploitGym.
 Author: grisun0
 License: GPL-3.0-or-later
@@ -282,6 +282,19 @@ Or install everything at once:
 pip install -e ".[train,lens,api,dev]"
 ```
 
+### Local fallback deployment (LazyOwn / TopoSwarm)
+
+Both repos ship an OpenAI-compatible fallback contract (full chat history,
+`tool` role, top-level `tools`, `best_of` reranking, `min_new_tokens`,
+deterministic `seed`, apt_chain tool validation). See
+[Fallback Integration](docs/FALLBACK_INTEGRATION.md); no code changes are
+required in LazyOwn or TopoSwarm, they only need the `base_url`:
+
+```bash
+bash scripts/serve_fallback.sh    # TopoGPT3 :8800/v1 + TopoExploit :8801/v1
+make eval-tasks                   # apt_chain JSONL task evals
+```
+
 ### Structured outputs and deterministic seeds
 
 The API server adopts the structured-output surface used by DeepInfer and

{topoexploit-0.2.0 → topoexploit-0.3.0}/README.md

@@ -239,6 +239,19 @@ Or install everything at once:
 pip install -e ".[train,lens,api,dev]"
 ```
 
+### Local fallback deployment (LazyOwn / TopoSwarm)
+
+Both repos ship an OpenAI-compatible fallback contract (full chat history,
+`tool` role, top-level `tools`, `best_of` reranking, `min_new_tokens`,
+deterministic `seed`, apt_chain tool validation). See
+[Fallback Integration](docs/FALLBACK_INTEGRATION.md); no code changes are
+required in LazyOwn or TopoSwarm, they only need the `base_url`:
+
+```bash
+bash scripts/serve_fallback.sh    # TopoGPT3 :8800/v1 + TopoExploit :8801/v1
+make eval-tasks                   # apt_chain JSONL task evals
+```
+
 ### Structured outputs and deterministic seeds
 
 The API server adopts the structured-output surface used by DeepInfer and

{topoexploit-0.2.0 → topoexploit-0.3.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "topoexploit"
-version = "0.2.0"
+version = "0.3.0"
 description = "TopoExploit: complex-valued spectral language model for security tasks (vulnerability analysis, patch analysis, exploit development), trained on ExploitGym."
 readme = "README.md"
 requires-python = ">=3.10"

topoexploit-0.3.0/tests/test_checkpoint_meta.py

@@ -0,0 +1,34 @@
+"""Checkpoint architecture discovery tests."""
+
+import pytest
+
+from topogpt3.checkpoint_meta import (
+    config_from_checkpoint,
+    probe_checkpoint,
+    read_state,
+)
+
+
+def test_read_state_missing(tmp_path):
+    assert read_state(tmp_path) == {}
+
+
+def test_probe_t3_small():
+    cfg, meta = config_from_checkpoint("checkpoints_topogpt3/last", device="cpu")
+    assert meta["scale"] == "small"
+    assert cfg.D_MODEL == 256 and cfg.N_LAYERS == 6
+    assert cfg.N_KV_HEADS == 2
+    assert cfg.MAX_SEQ_LEN == 512
+
+
+def test_probe_exploitgym_meta():
+    pytest.importorskip("safetensors")
+    meta = probe_checkpoint("checkpoints_topogpt3/last")
+    assert meta["vocab_size"] == 50257
+    assert meta["d_model"] == 256
+
+
+def test_config_matches_weights():
+    cfg, meta = config_from_checkpoint("checkpoints_topogpt3/last", device="cpu")
+    assert cfg.D_MODEL == meta["d_model"]
+    assert cfg.N_LAYERS == meta["n_layers"]

topoexploit-0.3.0/tests/test_data_pipeline.py

@@ -0,0 +1,61 @@
+"""Data pipeline tests: dedup, filters, teacher/verify/export (offline)."""
+
+import json
+
+from topogpt3.data_pipeline import (
+    dedup_stream,
+    normalize_text,
+    quality_filter,
+    to_chat_row,
+    verify_completion,
+    write_jsonl,
+)
+
+
+def test_normalize_strips_control_and_collapses_space():
+    text = "a \x00b\t\t c\n\n\n\n\n\nd"
+    cleaned = normalize_text(text)
+    assert "\x00" not in cleaned
+    assert "b c" in cleaned
+    assert cleaned.count("\n\n") == 1
+
+
+def test_dedup_exact_and_fuzzy():
+    rows = [
+        {"text": "hello world"},
+        {"text": "hello  world"},
+        {"text": "different content here"},
+    ]
+    out = list(dedup_stream(rows, fuzzy=False))
+    assert len(out) == 2
+    out = list(dedup_stream(rows, fuzzy=True))
+    assert len(out) == 2
+
+
+def test_quality_filter_bounds():
+    ok, reason = quality_filter("short")
+    assert not ok
+    ok, reason = quality_filter("x" * 600)
+    assert not ok and reason == "long_line"
+    ok, _reason = quality_filter("valid content with enough chars")
+    assert ok
+
+
+def test_verify_python_ast_and_sandbox():
+    ok, _ = verify_completion("```python\nprint(1)\n```", mode="python_ast")
+    assert ok
+    ok, _reason = verify_completion("```python\ndef broken(:\n```", mode="python_ast")
+    assert not ok
+
+
+def test_write_jsonl_roundtrip(tmp_path):
+    path = tmp_path / "out.jsonl"
+    n = write_jsonl(path, [{"a": 1}, {"a": 2}])
+    rows = [json.loads(line) for line in path.read_text().splitlines()]
+    assert n == 2 and rows[0]["a"] == 1
+
+
+def test_to_chat_row():
+    row = to_chat_row({"prompt": "p", "completion": "c"})
+    assert row["messages"][0]["role"] == "user"
+    assert row["messages"][1]["role"] == "assistant"

topoexploit-0.3.0/tests/test_eval_tasks.py

@@ -0,0 +1,42 @@
+"""Generic task eval tests (no model load)."""
+
+import json
+
+from eval.tasks import (
+    evaluate_cases,
+    load_cases,
+    summary,
+    verify_regex,
+    verify_tool_call,
+)
+from topogpt3.chat import TOOL_CALL_CLOSE, TOOL_CALL_OPEN
+from topogpt3.fallback import resolve_profile, synthesize_tool_call
+
+
+def _gen(prompt, profile_name=""):
+    call = synthesize_tool_call(resolve_profile(profile_name), prompt)
+    return f"{TOOL_CALL_OPEN}{json.dumps(call)}{TOOL_CALL_CLOSE}" if call else ""
+
+
+def test_apt_chain_cases_full_pass():
+    cases = load_cases("eval/apt_chain_cases.jsonl")
+    results = evaluate_cases(cases, _gen)
+    stats = summary(results)
+    assert stats["accuracy"] == 1.0, [
+        (r["id"], r["reason"]) for r in results if not r["ok"]
+    ]
+
+
+def test_regex_verifier():
+    case = {"verifier": {"type": "regex", "pattern": r"[A-Z]{2}-[0-9]{3}"}}
+    assert verify_regex(case, "AB-123")[0]
+    assert not verify_regex(case, "ab-123")[0]
+
+
+def test_tool_call_verifier_rejects_wrong_target():
+    case = {
+        "verifier": {"type": "apt_chain", "tool": "apt_chain", "target": "10.0.0.9"}
+    }
+    completion = _gen("Run the full chain against 10.10.11.78.", "apt-chain")
+    ok, reason = verify_tool_call(case, completion)
+    assert not ok and "target" in reason

topoexploit-0.3.0/tests/test_fallback.py

@@ -0,0 +1,133 @@
+"""Fallback profiles, apt_chain validation and candidate reranking tests."""
+
+import json
+
+from topogpt3.chat import TOOL_CALL_CLOSE, TOOL_CALL_OPEN
+from topogpt3.fallback import (
+    APT_CHAIN_TOOL,
+    LAZYOWN_TOOLS,
+    PROFILES,
+    rank_candidates,
+    resolve_profile,
+    score_candidate,
+    synthesize_tool_call,
+    tool_calls_payload,
+    validate_tool_call,
+)
+
+
+def _call(name, args):
+    return f"{TOOL_CALL_OPEN}{json.dumps({'name': name, 'arguments': args})}{TOOL_CALL_CLOSE}"
+
+
+class TestProfiles:
+    def test_resolve_by_name_and_alias(self):
+        assert resolve_profile("topogpt3").name == "topogpt3"
+        assert resolve_profile("apt_chain").name == "topoexploit-apt-chain"
+        assert resolve_profile(None).name == "topogpt3"
+
+    def test_unknown_falls_back(self):
+        assert resolve_profile("nope-xyz").name == "topogpt3"
+
+    def test_apt_chain_tools_registered(self):
+        names = {t["function"]["name"] for t in LAZYOWN_TOOLS}
+        assert "apt_chain" in names
+        assert APT_CHAIN_TOOL["function"]["parameters"]["required"] == ["target"]
+
+
+class TestValidateToolCall:
+    def test_valid_apt_chain(self):
+        ok, reason = validate_tool_call(
+            {
+                "name": "apt_chain",
+                "arguments": {"target": "10.10.11.78", "dry_run": True},
+            }
+        )
+        assert ok, reason
+
+    def test_requires_target(self):
+        ok, reason = validate_tool_call({"name": "apt_chain", "arguments": {}})
+        assert not ok and "target" in reason
+
+    def test_invalid_chain_name(self):
+        ok, _reason = validate_tool_call(
+            {
+                "name": "apt_chain",
+                "arguments": {"target": "10.0.0.1", "chain": "bad chain;"},
+            }
+        )
+        assert not ok
+
+    def test_run_command_forbidden_chars(self):
+        ok, _ = validate_tool_call(
+            {"name": "run_command", "arguments": {"command": "ls; rm -rf /"}}
+        )
+        assert not ok
+
+    def test_unknown_tool(self):
+        ok, reason = validate_tool_call({"name": "web_get", "arguments": {}})
+        assert not ok and "unknown" in reason
+
+
+class TestScoring:
+    def test_tool_profile_prefers_valid_call(self):
+        profile = resolve_profile("apt-chain")
+        good = _call("apt_chain", {"target": "10.10.11.78", "dry_run": True})
+        prose = "I would run the chain for you."
+        best_text, score = "", -1e18
+        for text in (prose, good):
+            s, _d = score_candidate(text, profile)
+            if s > score:
+                best_text, score = text, s
+        assert best_text == good and score > 0
+
+    def test_require_tool_penalizes_prose(self):
+        profile = resolve_profile("apt-chain")
+        score, _ = score_candidate("Sure, tell me more.", profile)
+        assert score < 0
+
+    def test_code_profile_scores_parseable_code(self):
+        profile = resolve_profile("topogpt3-code")
+        good = "```python\ndef f(n):\n    return n+1\n```"
+        bad = "```python\ndef f(n):\n    return n+\n```"
+        assert score_candidate(good, profile)[0] > score_candidate(bad, profile)[0]
+
+    def test_rank_candidates_sorted(self):
+        profile = resolve_profile("apt-chain")
+        ranked = rank_candidates(
+            ["prose", _call("apt_chain", {"target": "10.0.0.1"})], profile
+        )
+        assert ranked[0][2].find("apt_chain") != -1
+
+
+class TestSynthesize:
+    def test_ip_prompt(self):
+        profile = resolve_profile("apt-chain")
+        call = synthesize_tool_call(profile, "Run the full chain against 10.10.11.78.")
+        assert call and call["name"] == "apt_chain"
+        assert call["arguments"]["target"] == "10.10.11.78"
+        assert call["arguments"]["dry_run"] is True
+
+    def test_playbook_extraction(self):
+        call = synthesize_tool_call(
+            resolve_profile("apt-chain"),
+            "Emulate APT29 with chain apt_rocke against 10.0.0.5",
+        )
+        assert call and call["arguments"]["chain"] == "apt_rocke"
+
+    def test_no_target_no_call(self):
+        assert synthesize_tool_call(resolve_profile("apt-chain"), "hello world") is None
+
+
+class TestToolCallsPayload:
+    def test_openai_shape(self):
+        calls = tool_calls_payload(_call("apt_chain", {"target": "10.0.0.1"}))
+        assert calls[0]["type"] == "function"
+        assert calls[0]["function"]["name"] == "apt_chain"
+        args = json.loads(calls[0]["function"]["arguments"])
+        assert args["target"] == "10.0.0.1"
+
+
+def test_all_profiles_have_system():
+    for profile in PROFILES.values():
+        assert profile.system.strip()

{topoexploit-0.2.0 → topoexploit-0.3.0}/topoexploit.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: topoexploit
-Version: 0.2.0
+Version: 0.3.0
 Summary: TopoExploit: complex-valued spectral language model for security tasks (vulnerability analysis, patch analysis, exploit development), trained on ExploitGym.
 Author: grisun0
 License: GPL-3.0-or-later
@@ -282,6 +282,19 @@ Or install everything at once:
 pip install -e ".[train,lens,api,dev]"
 ```
 
+### Local fallback deployment (LazyOwn / TopoSwarm)
+
+Both repos ship an OpenAI-compatible fallback contract (full chat history,
+`tool` role, top-level `tools`, `best_of` reranking, `min_new_tokens`,
+deterministic `seed`, apt_chain tool validation). See
+[Fallback Integration](docs/FALLBACK_INTEGRATION.md); no code changes are
+required in LazyOwn or TopoSwarm, they only need the `base_url`:
+
+```bash
+bash scripts/serve_fallback.sh    # TopoGPT3 :8800/v1 + TopoExploit :8801/v1
+make eval-tasks                   # apt_chain JSONL task evals
+```
+
 ### Structured outputs and deterministic seeds
 
 The API server adopts the structured-output surface used by DeepInfer and

{topoexploit-0.2.0 → topoexploit-0.3.0}/topoexploit.egg-info/SOURCES.txt

@@ -1,6 +1,10 @@
 LICENSE
 README.md
 pyproject.toml
+tests/test_checkpoint_meta.py
+tests/test_data_pipeline.py
+tests/test_eval_tasks.py
+tests/test_fallback.py
 tests/test_jlens.py
 tests/test_lens_model.py
 tests/test_structured_output.py
@@ -13,10 +17,15 @@ topoexploit.egg-info/top_level.txt
 topogpt3/__init__.py
 topogpt3/__main__.py
 topogpt3/api_server.py
+topogpt3/chat.py
+topogpt3/checkpoint_meta.py
 topogpt3/continuation.py
+topogpt3/data_pipeline.py
+topogpt3/eval_toolcall.py
 topogpt3/ewc.py
 topogpt3/exploitgym_config.py
 topogpt3/exploitgym_loader.py
+topogpt3/fallback.py
 topogpt3/hodge_cm.py
 topogpt3/inference.py
 topogpt3/inference_hrm.py
@@ -25,5 +34,7 @@ topogpt3/lens_model.py
 topogpt3/merged_config.py
 topogpt3/model.py
 topogpt3/structured_output.py
+topogpt3/tokenizer_tools.py
+topogpt3/tools_agent.py
 topogpt3/train.py
 topogpt3/vanilla_control.py