File size: 8,206 Bytes
dfb775d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 | """Held-out CE-loss scorer for an adapted checkpoint.
The heavyweight torch/transformers/peft path is mocked here — real
end-to-end coverage happens via `mindxtrain eval-checkpoint` against
a freshly trained adapter in the `_real` recipe demo run. These tests
just verify the orchestration (JSONL iteration, sink callback, both
loss invocations, summary shape) is sound.
"""
from __future__ import annotations
import json
import sys
import types
from pathlib import Path
import pytest
def _write_jsonl(path: Path, rows: list[dict]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("w", encoding="utf-8") as fh:
for row in rows:
fh.write(json.dumps(row) + "\n")
def _install_fake_ml_stack(monkeypatch, base_loss: float, adapter_loss: float) -> None:
"""Install fake `transformers` + `peft` + `torch` modules.
The fake AutoModelForCausalLM returns sequentially `base_loss` then
`adapter_loss` so the score function picks up a non-trivial delta.
"""
class _FakeOutput:
def __init__(self, loss_value: float) -> None:
class _T:
def detach(self_inner):
return self_inner
def cpu(self_inner):
return self_inner
def item(self_inner):
return loss_value
self.loss = _T()
class _FakeBaseModel:
# Base-model branch — returns base_loss every call.
_loss = base_loss
def eval(self):
return self
def __call__(self, **_kw):
return _FakeOutput(self.__class__._loss)
class _FakeTokenizer:
def apply_chat_template(self, messages, tokenize=False, add_generation_prompt=False):
return " ".join(m.get("content", "") for m in messages)
def __call__(self, text, return_tensors="pt", truncation=True, max_length=1024):
import torch # the fake one
return {"input_ids": torch.tensor([[0, 1, 2]])}
class _FakeAutoModel:
@staticmethod
def from_pretrained(name, *a, **kw):
return _FakeBaseModel()
class _FakeAutoTokenizer:
@staticmethod
def from_pretrained(name, *a, **kw):
return _FakeTokenizer()
fake_transformers = types.ModuleType("transformers")
fake_transformers.AutoModelForCausalLM = _FakeAutoModel
fake_transformers.AutoTokenizer = _FakeAutoTokenizer
monkeypatch.setitem(sys.modules, "transformers", fake_transformers)
class _FakePeftModel(_FakeBaseModel):
# Adapter branch — overrides _loss so PeftModel calls return adapter_loss.
_loss = adapter_loss
class _FakePeftFactory:
@staticmethod
def from_pretrained(base, adapter_dir):
return _FakePeftModel()
fake_peft = types.ModuleType("peft")
fake_peft.PeftModel = _FakePeftFactory
monkeypatch.setitem(sys.modules, "peft", fake_peft)
class _FakeTorch:
@staticmethod
def tensor(data):
return data
class no_grad:
def __enter__(self): return self
def __exit__(self, *a): return False
fake_torch = types.ModuleType("torch")
fake_torch.tensor = _FakeTorch.tensor
fake_torch.no_grad = _FakeTorch.no_grad
monkeypatch.setitem(sys.modules, "torch", fake_torch)
def test_score_checkpoint_returns_summary(tmp_path, monkeypatch):
"""Happy path: 2 rows, mocked models, delta == adapter - base."""
_install_fake_ml_stack(monkeypatch, base_loss=2.5, adapter_loss=2.1)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "training.jsonl"
_write_jsonl(jsonl, [
{"messages": [{"role": "user", "content": "what is mindX?"},
{"role": "assistant", "content": "a cognitive runtime"}]},
{"messages": [{"role": "user", "content": "dream cycle?"},
{"role": "assistant", "content": "phase 5b output"}]},
])
score = score_checkpoint(
adapter_dir=tmp_path / "adapter",
base_model="HuggingFaceTB/SmolLM2-135M",
jsonl_path=jsonl,
max_samples=8,
)
assert score.n == 2
assert score.base_loss == pytest.approx(2.5)
assert score.adapter_loss == pytest.approx(2.1)
assert score.delta == pytest.approx(-0.4)
assert score.delta < 0
def test_score_checkpoint_streams_through_sink(tmp_path, monkeypatch):
_install_fake_ml_stack(monkeypatch, base_loss=2.0, adapter_loss=2.0)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "t.jsonl"
_write_jsonl(jsonl, [{"messages": [{"role": "user", "content": "x"}]}])
lines: list[str] = []
score_checkpoint(
adapter_dir=tmp_path / "adapter",
base_model="b",
jsonl_path=jsonl,
max_samples=1,
sink=lines.append,
)
joined = "\n".join(lines)
# The progress prints surface key stages so the CLI / Coach UI can
# show the user what's happening during a multi-minute scoring run.
assert "held-out rows from" in joined
assert "base mean loss" in joined
assert "adapter mean loss" in joined
assert "delta" in joined
def test_score_checkpoint_empty_jsonl_raises(tmp_path, monkeypatch):
_install_fake_ml_stack(monkeypatch, 2.0, 2.0)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "empty.jsonl"
jsonl.write_text("", encoding="utf-8")
with pytest.raises(ValueError, match="no readable JSONL rows"):
score_checkpoint(
adapter_dir=tmp_path / "adapter",
base_model="b",
jsonl_path=jsonl,
)
def test_score_checkpoint_skips_malformed_lines(tmp_path, monkeypatch):
"""Bad JSON / non-messages rows are silently skipped (matches iter_mindx_dreams)."""
_install_fake_ml_stack(monkeypatch, 1.5, 1.2)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "messy.jsonl"
jsonl.write_text(
"not-json\n"
'{"no_messages": true}\n'
'{"messages": [{"role": "user", "content": "real row"}]}\n',
encoding="utf-8",
)
score = score_checkpoint(
adapter_dir=tmp_path / "adapter", base_model="b", jsonl_path=jsonl,
)
assert score.n == 1
def test_score_checkpoint_caps_at_max_samples(tmp_path, monkeypatch):
_install_fake_ml_stack(monkeypatch, 1.0, 0.9)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "big.jsonl"
_write_jsonl(jsonl, [
{"messages": [{"role": "user", "content": f"row {i}"}]}
for i in range(50)
])
score = score_checkpoint(
adapter_dir=tmp_path / "adapter", base_model="b", jsonl_path=jsonl,
max_samples=3,
)
assert score.n == 3
def test_score_checkpoint_missing_extras_raises(tmp_path, monkeypatch):
"""No peft/transformers → ImportError with the install hint."""
# Block the imports inside score_checkpoint by removing them from
# sys.modules and shadowing the importer.
monkeypatch.setitem(sys.modules, "peft", None)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "t.jsonl"
_write_jsonl(jsonl, [{"messages": [{"role": "user", "content": "x"}]}])
with pytest.raises(ImportError, match="uv sync --extra ml"):
score_checkpoint(
adapter_dir=tmp_path / "adapter", base_model="b", jsonl_path=jsonl,
)
def test_as_dict_round_trip(tmp_path, monkeypatch):
_install_fake_ml_stack(monkeypatch, 2.0, 1.7)
from mindxtrain.eval.held_out_loss import score_checkpoint
jsonl = tmp_path / "t.jsonl"
_write_jsonl(jsonl, [{"messages": [{"role": "user", "content": "x"}]}])
score = score_checkpoint(
adapter_dir=tmp_path / "adapter", base_model="b", jsonl_path=jsonl,
)
d = score.as_dict()
# The CLI uses console.print_json(score.as_dict()), so it must be
# a flat dict with the documented keys.
assert set(d) == {"base_model", "adapter_dir", "n", "base_loss", "adapter_loss", "delta"}
assert d["base_loss"] == pytest.approx(2.0)
assert d["adapter_loss"] == pytest.approx(1.7)
|