Skip to content

Personal AI, On Personal Devices

Python
5.9 KB
Raw Blame
"""Tests for orchestrator SFT trainer."""

from __future__ import annotations

from unittest.mock import MagicMock

import pytest

from openjarvis.learning.intelligence.orchestrator.sft_trainer import (
    OrchestratorSFTConfig,
    OrchestratorSFTDataset,
)


class TestOrchestratorSFTConfig:
    def test_defaults(self):
        cfg = OrchestratorSFTConfig()
        assert cfg.model_name == "Qwen/Qwen3-1.7B"
        assert cfg.num_epochs == 3
        assert cfg.batch_size == 8
        assert cfg.learning_rate == 2e-5
        assert cfg.max_seq_length == 4096
        assert cfg.gradient_checkpointing is True

    def test_custom_values(self):
        cfg = OrchestratorSFTConfig(
            model_name="test-model",
            num_epochs=5,
            batch_size=16,
        )
        assert cfg.model_name == "test-model"
        assert cfg.num_epochs == 5
        assert cfg.batch_size == 16

    def test_default_tools(self):
        cfg = OrchestratorSFTConfig()
        assert "calculator" in cfg.available_tools
        assert "think" in cfg.available_tools


class TestOrchestratorSFTDataset:
    def test_empty_on_missing_file(self):
        tok = MagicMock()
        ds = OrchestratorSFTDataset(
            trace_path="/nonexistent/path.jsonl",
            tokenizer=tok,
        )
        assert len(ds) == 0

    def test_format_conversation_fallback(self, tmp_path):
        """Test manual formatting when tokenizer has no chat template."""
        import json

        trace_file = tmp_path / "traces.jsonl"
        trace = {
            "conversations": [
                {"role": "user", "content": "Hello"},
                {"role": "assistant", "content": "Hi there"},
            ],
        }
        trace_file.write_text(json.dumps(trace) + "\n")

        tok = MagicMock()
        tok.eos_token = "</s>"
        del tok.apply_chat_template  # no chat template

        ds = OrchestratorSFTDataset(
            trace_path=str(trace_file),
            tokenizer=tok,
        )
        assert len(ds) == 1

        text = ds._format_conversation(trace["conversations"])
        assert "<|user|>" in text
        assert "Hello" in text
        assert "<|assistant|>" in text
        assert "Hi there" in text
        assert text.endswith("</s>")

    def test_format_tool_message(self):
        tok = MagicMock()
        tok.eos_token = ""
        del tok.apply_chat_template

        ds = OrchestratorSFTDataset(
            trace_path="/nonexistent",
            tokenizer=tok,
        )
        convs = [
            {"role": "tool", "name": "calculator", "content": "42"},
        ]
        text = ds._format_conversation(convs)
        assert "calculator" in text
        assert "42" in text

    def test_iter_batches(self, tmp_path):
        import json

        trace_file = tmp_path / "traces.jsonl"
        traces = []
        for i in range(5):
            traces.append(
                {
                    "conversations": [
                        {"role": "user", "content": f"q{i}"},
                        {"role": "assistant", "content": f"a{i}"},
                    ]
                }
            )
        trace_file.write_text("\n".join(json.dumps(t) for t in traces) + "\n")

        tok = MagicMock()
        tok.eos_token = ""
        del tok.apply_chat_template
        tok.return_value = {
            "input_ids": MagicMock(),
            "attention_mask": MagicMock(),
        }

        ds = OrchestratorSFTDataset(
            trace_path=str(trace_file),
            tokenizer=tok,
        )
        batches = list(ds.iter_batches(batch_size=2))
        assert len(batches) == 3  # 2+2+1


class TestSFTLabelMasking:
    """Regression for #521: padding positions must be excluded from the SFT loss.

    With ``padding="max_length"`` and ``pad_token == eos_token``, an unmasked
    ``labels`` makes the model optimise "predict EOS at a padded position" for
    the bulk of every example, diluting the gradient on real content and
    understating the reported loss. ``labels`` must be ``-100`` wherever
    ``attention_mask == 0`` and equal to ``input_ids`` elsewhere — without
    mutating ``input_ids`` (the pre-fix code aliased the two).
    """

    def test_getitem_masks_padding_positions(self, tmp_path):
        torch = pytest.importorskip("torch")
        import json

        trace_file = tmp_path / "traces.jsonl"
        trace_file.write_text(
            json.dumps(
                {
                    "conversations": [
                        {"role": "user", "content": "hi"},
                        {"role": "assistant", "content": "yo"},
                    ]
                }
            )
            + "\n"
        )

        class _FakeTokenizer:
            """Returns a fixed padded encoding: 2 real tokens, 6 pad (id 0)."""

            eos_token = "</s>"

            def __call__(self, text, **kwargs):
                input_ids = torch.tensor([[11, 12, 0, 0, 0, 0, 0, 0]])
                attention_mask = torch.tensor([[1, 1, 0, 0, 0, 0, 0, 0]])
                return {"input_ids": input_ids, "attention_mask": attention_mask}

        ds = OrchestratorSFTDataset(
            trace_path=str(trace_file), tokenizer=_FakeTokenizer()
        )
        item = ds[0]
        ids, mask, labels = item["input_ids"], item["attention_mask"], item["labels"]

        assert (labels[mask == 0] == -100).all()  # padded -> ignored by loss
        assert (labels[mask == 1] == ids[mask == 1]).all()  # real -> unchanged
        assert (ids[mask == 0] != -100).all()  # input_ids not mutated in place
        assert not torch.equal(ids, labels)  # masked clone, not an alias


class TestSFTRegistration:
    def test_registered_in_learning_registry(self):
        # Import to trigger registration
        import openjarvis.learning.intelligence.orchestrator.sft_trainer  # noqa: F401
        from openjarvis.core.registry import LearningRegistry

        assert LearningRegistry.contains("orchestrator_sft")