@@ -0,0 +1,125 @@ |
| 1 | +"""Integration test: tool_use_fidelity end-to-end on a real tiny adapter. |
| 2 | + |
| 3 | +Builds a tiny random LoRA on SmolLM2-135M-Instruct and runs the probe |
| 4 | +against a single tool-use case. The intent isn't to assert that the |
| 5 | +135M base produces useful tool calls — it almost certainly won't — |
| 6 | +but to exercise the full code path on a real backend so a regression |
| 7 | +in the JSON-extraction / schema-check / hallucination plumbing |
| 8 | +surfaces in slow CI rather than only at user-fix time. |
| 9 | + |
| 10 | +Marked ``slow + online``. |
| 11 | +""" |
| 12 | + |
| 13 | +from __future__ import annotations |
| 14 | + |
| 15 | +from pathlib import Path |
| 16 | + |
| 17 | +import pytest |
| 18 | + |
| 19 | +pytestmark = [pytest.mark.slow, pytest.mark.online] |
| 20 | + |
| 21 | + |
| 22 | +def _build_random_lora_adapter(base_dir: Path, out_dir: Path) -> None: |
| 23 | + """A trivially-random LoRA over q_proj/v_proj — same shape used by |
| 24 | + the other slow-lane backend tests.""" |
| 25 | + import torch |
| 26 | + from peft import LoraConfig, get_peft_model |
| 27 | + from transformers import AutoModelForCausalLM, AutoTokenizer |
| 28 | + |
| 29 | + torch.manual_seed(0) |
| 30 | + tokenizer = AutoTokenizer.from_pretrained(str(base_dir)) |
| 31 | + if tokenizer.pad_token_id is None: |
| 32 | + tokenizer.pad_token = tokenizer.eos_token |
| 33 | + base = AutoModelForCausalLM.from_pretrained(str(base_dir), torch_dtype=torch.float32) |
| 34 | + cfg = LoraConfig( |
| 35 | + r=8, |
| 36 | + lora_alpha=16, |
| 37 | + target_modules=["q_proj", "v_proj"], |
| 38 | + lora_dropout=0.0, |
| 39 | + bias="none", |
| 40 | + task_type="CAUSAL_LM", |
| 41 | + ) |
| 42 | + peft_model = get_peft_model(base, cfg) |
| 43 | + with torch.no_grad(): |
| 44 | + for name, param in peft_model.named_parameters(): |
| 45 | + if "lora_B" in name: |
| 46 | + # Tiny perturbation — enough that base.generate ≠ ft.generate |
| 47 | + # but small enough to keep generations finite + sensible. |
| 48 | + param.copy_(torch.randn_like(param) * 0.01) |
| 49 | + peft_model.save_pretrained(str(out_dir)) |
| 50 | + tokenizer.save_pretrained(str(out_dir)) |
| 51 | + |
| 52 | + |
| 53 | +@pytest.fixture(scope="module") |
| 54 | +def random_adapter(tiny_model_dir: Path, tmp_path_factory: pytest.TempPathFactory) -> Path: |
| 55 | + out = tmp_path_factory.mktemp("tool-use-fidelity-adapter") |
| 56 | + _build_random_lora_adapter(tiny_model_dir, out) |
| 57 | + return out |
| 58 | + |
| 59 | + |
| 60 | +def test_probe_runs_end_to_end_on_real_adapter(tiny_model_dir: Path, random_adapter: Path) -> None: |
| 61 | + """Smoke: HF backend + real adapter + probe execution returns a finalized |
| 62 | + result with all evidence keys populated. SmolLM2-135M can't reliably |
| 63 | + emit OpenAI-shape calls, so the probe will likely FAIL the validity |
| 64 | + floor — what we assert here is that it produced a structured verdict |
| 65 | + + finite metrics, not a particular pass/fail outcome.""" |
| 66 | + from dlm_sway.backends.hf import HuggingFaceDifferentialBackend |
| 67 | + from dlm_sway.core.model import ModelSpec |
| 68 | + from dlm_sway.core.result import Verdict |
| 69 | + from dlm_sway.probes.base import RunContext, build_probe |
| 70 | + |
| 71 | + backend = HuggingFaceDifferentialBackend( |
| 72 | + base_spec=ModelSpec(base=str(tiny_model_dir), kind="hf", dtype="fp32", device="cpu"), |
| 73 | + adapter_path=random_adapter, |
| 74 | + ) |
| 75 | + try: |
| 76 | + probe, spec = build_probe( |
| 77 | + { |
| 78 | + "name": "tuf_smoke", |
| 79 | + "kind": "tool_use_fidelity", |
| 80 | + "cases": [ |
| 81 | + { |
| 82 | + "prompt": ( |
| 83 | + "You are a tool-using assistant. The user asks: " |
| 84 | + "search the web for cats. Reply with ONLY a JSON " |
| 85 | + 'object of the form {"name": ..., "arguments": {...}}.' |
| 86 | + ), |
| 87 | + "tool_spec": { |
| 88 | + "name": "search_web", |
| 89 | + "parameters": { |
| 90 | + "type": "object", |
| 91 | + "properties": { |
| 92 | + "query": {"type": "string"}, |
| 93 | + "max_results": {"type": "integer"}, |
| 94 | + }, |
| 95 | + "required": ["query"], |
| 96 | + }, |
| 97 | + }, |
| 98 | + "gold_tool_name": "search_web", |
| 99 | + "max_new_tokens": 64, |
| 100 | + } |
| 101 | + ], |
| 102 | + "allowed_tools": ["search_web"], |
| 103 | + } |
| 104 | + ) |
| 105 | + ctx = RunContext(backend=backend, seed=0) |
| 106 | + result = probe.run(spec, ctx) |
| 107 | + finally: |
| 108 | + backend.close() |
| 109 | + |
| 110 | + # End-to-end shape: a verdict came back, evidence carries every |
| 111 | + # documented key, and the rates are in [0, 1]. |
| 112 | + assert result.verdict in {Verdict.PASS, Verdict.FAIL, Verdict.WARN}, result.message |
| 113 | + assert result.evidence["num_cases"] == 1 |
| 114 | + for key in ( |
| 115 | + "json_valid_rate_base", |
| 116 | + "json_valid_rate_ft", |
| 117 | + "validity_delta", |
| 118 | + "mean_arg_disagreement", |
| 119 | + "hallucination_rate", |
| 120 | + ): |
| 121 | + assert key in result.evidence, f"missing evidence key {key}" |
| 122 | + assert 0.0 <= result.evidence["json_valid_rate_ft"] <= 1.0 |
| 123 | + assert 0.0 <= result.evidence["json_valid_rate_base"] <= 1.0 |
| 124 | + assert -1.0 <= result.evidence["validity_delta"] <= 1.0 |
| 125 | + assert 0.0 <= result.evidence["hallucination_rate"] <= 1.0 |