fix(loader): fallback field mapping for sonnet comparison schema

[nvandessel]

Summary

• Add fallback field mapping in JsonlLoader._parse_entry(): response ← sonnet_response, parsed ← haiku_parsed
• Sonnet comparison data (682 records) now loads correctly, yielding 2,076 total SFT pairs
• 4 new tests covering fallback behavior and preference for canonical fields

Test plan

• 4 new tests for sonnet comparison schema fallback (all pass)
• All 78 existing tests still pass
• Ruff linting clean
• Verified on real data: 1,394 haiku + 682 sonnet = 2,076 SFT pairs

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR adds fallback field mapping in JsonlLoader._parse_entry() to handle a sonnet comparison JSONL schema that stores responses and parsed output under sonnet_response / haiku_parsed instead of the canonical response / parsed keys. The fix enables 682 previously unloadable records to flow through the full pipeline (load → clean → format), increasing total SFT pairs from ~1,394 to ~2,076. A new Kaggle training notebook is also introduced.

Key changes:

• response falls back to sonnet_response and parsed falls back to haiku_parsed using consistent is not None guards — the concern raised in the previous review thread about or vs is not None asymmetry has been resolved; both fields now use the more precise is not None check.
• The cross-model pairing (sonnet_response + haiku_parsed) is semantically safe: parsed is not consumed by SftFormatter or DecisionCleaner — only response is used as the assistant training target, so the fields not being from the same model has no downstream effect.
• 4 new tests cover absent-key fallback, explicit-preference, and full pipeline scenarios.
• The notebook final summary cell references report without a fallback guard — a minor style issue if cells are run out of order.

Confidence Score: 5/5

Safe to merge — the loader fix is correct, consistently guarded, and well-tested; no P0/P1 findings.

All remaining findings are P2 (notebook style). The core loader change uses consistent is not None guards for all fallback fields, the parsed cross-model asymmetry is harmless since parsed is not consumed downstream in the training pipeline, and the 4 new tests provide solid coverage of the fallback paths.

notebooks/train-hippofloop.ipynb — minor guard missing for report in summary cell.

Important Files Changed

Filename	Overview
src/hippofloop/data/loader.py	Adds is not None fallback for response→sonnet_response and parsed→haiku_parsed; consistent guard strategy across all three fallback fields.
tests/data/test_loader.py	4 new tests covering fallback activation, canonical-field preference, and end-to-end pipeline; adequate coverage for the added logic.
notebooks/train-hippofloop.ipynb	New end-to-end training notebook for Kaggle; minor issue where the final summary cell references report without a guard, which would NameError if the evaluation cell is skipped.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Raw JSONL line] --> B{response key present\nand not None?}
    B -- Yes --> C[use response]
    B -- No --> D{sonnet_response\npresent?}
    D -- Yes --> E[use sonnet_response]
    D -- No --> F[use empty string]
    C --> G[DecisionEntry.response]
    E --> G
    F --> G

    A --> H{parsed key present\nand not None?}
    H -- Yes --> I[use parsed]
    H -- No --> J{haiku_parsed\npresent?}
    J -- Yes --> K[use haiku_parsed]
    J -- No --> L[use None]
    I --> M[DecisionEntry.parsed]
    K --> M
    L --> M

    G --> N[DecisionCleaner\nfilters on response]
    N --> O[SftFormatter\nuses response as\nassistant message]
    M --> P[metadata only\nnot used in training]

Loading

_{Reviews (4): Last reviewed commit: "refactor: eliminate double dict lookups ..." | Re-trigger Greptile}