[Neonkraft]

Summary

This PR introduces transforms required to adapt the tool call rows from the Dolci-Instruct dataset. Without it, the tool call rows are rendered incorrectly.

Fixes #63

Type of change

• Bug fix
• New feature
• Refactor
• Performance
• Documentation
• Maintenance

Validation

New tests are written and a small run (~20 steps) was tried on LUMI.

[Copilot]

🟡 Changes recommended

Unresolved parser defects can alter tool arguments, discard convertible rows, or abort dataset loading.

4 open issues

• Catch TypeError from literal_eval during guarded evaluation · New
• Avoid rewriting keyword text inside quoted string values · New
• Preserve source location on replacement JSON Constant nodes · New
• Ignore parentheses inside strings when splitting fallback calls · New

What changed in this PR

This PR adds a dolci_tool_use transform to preserve Dolci-Instruct-SFT tool data in the standard SFT pipeline, addressing the Qwen3 rendering loss described in #63.

Changes:

• Converts Dolci tool definitions, calls, and environment responses into standard tool fields.
• Preserves tool_calls through the SFT dataset schema.
• Adds parsing, rejection, and Qwen3 rendering tests.

File	Description
tests/​test_transforms.py	Tests Dolci conversion and tool-data preservation through rendering.
src/​post_training/​methods/​sft.py	Adds tool-call fields to the messages schema.
src/​post_training/​data/​transforms.py	Implements Dolci tool-use parsing and conversion.

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Unresolved evaluation safety, argument corruption, and template compatibility issues can disrupt preprocessing or produce incorrect training data.

4 open issues

• Reject subscript assignment targets in dataset expressions · New
• Unbounded dataset expressions can exhaust memory or stall workers · New
• Normalize string arguments for the olmo3-instruct-sft template · New
• Avoid repairing escaped names inside strings and comments · New

2 previously missed issues

In code that hasn't changed since last review

Prevent keyword-rename collisions and preserve original names

src/​post_training/​data/​transforms.py:701

A valid argument named _kw_from is not rewritten by _rename_keyword_args, but these lines still rename it to from. A call containing both from="EUR" and _kw_from="USD" also collapses the two arguments into one, silently losing a value. Use a collision-free rename mapping and restore only argument names that were actually rewritten.

Reject non-finite values during JSON round-trip validation

src/​post_training/​data/​transforms.py:703

The round-trip check still accepts infinity: wx.get(value=1e309) produces {"value": Infinity}, and Python's permissive json.loads returns the same infinite float. Qwen3 inserts this string verbatim, so the training target contains an invalid JSON tool call. Set allow_nan=False so non-finite values follow the existing rejection path.

4 resolved since last review

• Avoid rewriting keyword text inside quoted string values
• Catch TypeError from literal_eval during guarded evaluation
• Ignore parentheses inside strings when splitting fallback calls
• Preserve source location on replacement JSON Constant nodes

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Unresolved allocation-limit, dataset-mixing, and conversion correctness issues can disrupt preprocessing or corrupt tool-call training data.

3 open issues

• Bound string-format allocations before evaluation · New
• Align schemas when mixing native and transformed messages · New
• Normalize string arguments for the olmo3-instruct-sft template

4 previously missed issues

In code that hasn't changed since last review

Restore only keyword names actually rewritten

src/​post_training/​data/​transforms.py:761

This strips _kw_ from legitimate parameter names, not just names introduced by _rename_keyword_args. For example, wx.get(_kw_from="EUR") is accepted but becomes arguments {"from": "EUR"}, changing the tool invocation and potentially colliding with a separate from parameter. Track which argument names were actually rewritten, using collision-free temporary names, and restore only those.

Reject non-finite values during round-trip validation

src/​post_training/​data/​transforms.py:763

The round-trip check accepts infinity: wx.get(q=1e309) serializes to {"q": Infinity}, which Python's JSON decoder reads back as the same float. Qwen3 inserts this string verbatim, so the row trains an invalid JSON tool call instead of being rejected. Set allow_nan=False so non-finite values follow the existing ValueError rejection path.

Preserve valid tool names during carriage-return repair

src/​post_training/​data/​transforms.py:786

A bare carriage-return separator is treated as a lost initial r: wx.get(q='a')\rwx.get(q='b') becomes calls to wx.get and rwx.get. This silently changes the tool invocation and retains the unknown name. Consult the declared tool names before inserting a letter, preserving names that are already valid.

Restrict newline repair to separators outside strings

src/​post_training/​data/​transforms.py:811

When a block contains both a literal \n separator and a string argument containing )\n, the original block fails tokenization and this global substitution rewrites both occurrences. For example, the raw block wx.get(q="a)\nb")\nwx.get(q="b") gains a real newline inside the quoted argument, so the entire otherwise-convertible row is dropped. Restrict this repair to separators outside string literals, including when the original block does not tokenize successfully.

3 resolved since last review

• Unbounded dataset expressions can exhaust memory or stall workers
• Reject subscript assignment targets in dataset expressions
• Avoid repairing escaped names inside strings and comments

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Dataset expressions can bypass the allocation guard and exhaust preprocessing worker memory.

3 open issues

• Prevent indirect callees from bypassing allocation guards · New
• Align schemas when mixing native and transformed messages
• Normalize string arguments for the olmo3-instruct-sft template

1 resolved since last review

• Bound string-format allocations before evaluation

_{Review effort: Balanced}