| 1 |
"""`TrainingSummary` — post-run report serialized to `logs/train-*.summary.json`. |
| 2 |
|
| 3 |
One summary per training run. Captures "how did it go?" in a compact |
| 4 |
form that's cheap to read from the CLI (no torch / HF imports needed) |
| 5 |
and human-inspectable. The manifest's `training_runs` list links out |
| 6 |
to the summary file by path so `dlm show` can load details on demand. |
| 7 |
""" |
| 8 |
|
| 9 |
from __future__ import annotations |
| 10 |
|
| 11 |
import json |
| 12 |
from collections.abc import Iterable |
| 13 |
from dataclasses import dataclass |
| 14 |
from pathlib import Path |
| 15 |
|
| 16 |
from pydantic import BaseModel, ConfigDict, Field |
| 17 |
|
| 18 |
from dlm.io.atomic import write_text |
| 19 |
|
| 20 |
|
| 21 |
class ProbeOutput(BaseModel): |
| 22 |
"""One probe's prompt + generated response + optional reference.""" |
| 23 |
|
| 24 |
model_config = ConfigDict(extra="forbid", frozen=True) |
| 25 |
|
| 26 |
prompt: str |
| 27 |
response: str |
| 28 |
reference: str | None = None |
| 29 |
section_id: str = "" |
| 30 |
|
| 31 |
|
| 32 |
class SourceProvenanceRecord(BaseModel): |
| 33 |
"""Serialized per-directive ingestion bookkeeping. |
| 34 |
|
| 35 |
Mirrors `dlm.directives.SourceProvenance` as a pydantic model so |
| 36 |
the training summary JSON is self-describing. One record per |
| 37 |
`training.sources` entry; `file_count == 0` indicates a directive |
| 38 |
that matched nothing (worth flagging in the CLI). |
| 39 |
""" |
| 40 |
|
| 41 |
model_config = ConfigDict(extra="forbid", frozen=True) |
| 42 |
|
| 43 |
path: str |
| 44 |
file_count: int = Field(0, ge=0) |
| 45 |
total_bytes: int = Field(0, ge=0) |
| 46 |
skipped_binary: int = Field(0, ge=0) |
| 47 |
skipped_encoding: int = Field(0, ge=0) |
| 48 |
skipped_over_size: int = Field(0, ge=0) |
| 49 |
|
| 50 |
|
| 51 |
class TrainingSummary(BaseModel): |
| 52 |
"""Canonical post-run report.""" |
| 53 |
|
| 54 |
model_config = ConfigDict(extra="forbid", frozen=True) |
| 55 |
|
| 56 |
run_id: int = Field(..., ge=1) |
| 57 |
adapter_version: int = Field(..., ge=1) |
| 58 |
seed: int |
| 59 |
final_train_loss: float | None = None |
| 60 |
final_val_loss: float | None = None |
| 61 |
final_val_perplexity: float | None = None |
| 62 |
retention_loss: float | None = None |
| 63 |
retention_loss_delta: float | None = None |
| 64 |
# Mixed-mode breakdown: when the training rows mix CPT prose and |
| 65 |
# SFT instruction rows, the aggregate loss hides which side is |
| 66 |
# driving movement. These fields split the final train/val and |
| 67 |
# retention metrics by row mode. All optional — single-mode runs |
| 68 |
# leave them None. |
| 69 |
train_loss_cpt: float | None = None |
| 70 |
train_loss_sft: float | None = None |
| 71 |
val_loss_cpt: float | None = None |
| 72 |
val_loss_sft: float | None = None |
| 73 |
retention_cpt: float | None = None |
| 74 |
retention_sft: float | None = None |
| 75 |
probes: list[ProbeOutput] = Field(default_factory=list) |
| 76 |
early_stopped: bool = False |
| 77 |
steps: int = Field(0, ge=0) |
| 78 |
duration_seconds: float = Field(0.0, ge=0.0) |
| 79 |
determinism_class: str = "best_effort" |
| 80 |
# Per-directive ingestion provenance. Empty when no |
| 81 |
# `training.sources` declared. Order matches the frontmatter so |
| 82 |
# CLI formatters can line up rows with source entries. |
| 83 |
source_directives: list[SourceProvenanceRecord] = Field(default_factory=list) |
| 84 |
|
| 85 |
|
| 86 |
@dataclass(frozen=True) |
| 87 |
class LossByMode: |
| 88 |
"""Mean loss split by row mode (`cpt` prose vs `sft` instruction). |
| 89 |
|
| 90 |
`None` where the corresponding row count was zero — the caller |
| 91 |
stores that verbatim in the summary so ``None`` is honest "we had |
| 92 |
no rows of that mode" rather than a zeroed-out number. |
| 93 |
""" |
| 94 |
|
| 95 |
cpt: float | None |
| 96 |
sft: float | None |
| 97 |
|
| 98 |
|
| 99 |
def split_loss_by_mode(rows: Iterable[tuple[float, str]]) -> LossByMode: |
| 100 |
"""Average `(loss, mode)` pairs grouped by mode. |
| 101 |
|
| 102 |
`mode` is expected to be one of `"cpt"` or `"sft"`; other strings |
| 103 |
are ignored so the caller can pass a single stream containing |
| 104 |
preference/other rows without pre-filtering. |
| 105 |
""" |
| 106 |
cpt_losses: list[float] = [] |
| 107 |
sft_losses: list[float] = [] |
| 108 |
for loss, mode in rows: |
| 109 |
if mode == "cpt": |
| 110 |
cpt_losses.append(loss) |
| 111 |
elif mode == "sft": |
| 112 |
sft_losses.append(loss) |
| 113 |
return LossByMode( |
| 114 |
cpt=sum(cpt_losses) / len(cpt_losses) if cpt_losses else None, |
| 115 |
sft=sum(sft_losses) / len(sft_losses) if sft_losses else None, |
| 116 |
) |
| 117 |
|
| 118 |
|
| 119 |
def save_summary(path: Path, summary: TrainingSummary) -> None: |
| 120 |
"""Atomically serialize `summary` as pretty JSON. |
| 121 |
|
| 122 |
Uses the atomic-write helper so a concurrent CLI reader never sees |
| 123 |
a torn file. |
| 124 |
""" |
| 125 |
payload = summary.model_dump(mode="json") |
| 126 |
blob = json.dumps(payload, sort_keys=True, indent=2) + "\n" |
| 127 |
write_text(path, blob) |
| 128 |
|
| 129 |
|
| 130 |
def load_summary(path: Path) -> TrainingSummary: |
| 131 |
"""Inverse of `save_summary`; raises `pydantic.ValidationError` on drift.""" |
| 132 |
raw = path.read_text(encoding="utf-8") |
| 133 |
data = json.loads(raw) |
| 134 |
return TrainingSummary.model_validate(data) |
| 135 |
|
| 136 |
|
| 137 |
def summary_path_for(logs_dir: Path, run_id: int, started_iso: str) -> Path: |
| 138 |
"""Match the JSONL log file naming so pairs are easy to eyeball. |
| 139 |
|
| 140 |
`train-<run_id>-<ts>.summary.json` sits next to the `.jsonl` log |
| 141 |
with the same stem. |
| 142 |
""" |
| 143 |
ts = started_iso.replace(":", "").replace("-", "") |
| 144 |
return logs_dir / f"train-{run_id:06d}-{ts}.summary.json" |