[yudigege86]

Motivation

Muse Spark renders ShareGPT rows through its own tokenizer as Harmony messages (<|start|>assistant to=self<|message|>...<|eom|> and to=user closed by <|eot|>). data.chat_template: qwen3.5 only selects a loss-mask parser, and that parser searches for <|im_start|>assistant markers this template never emits. Every mask is empty and the rows are dropped.

This adds a muse profile that keeps rendering on tokenizer.apply_chat_template and masks the assistant spans that template actually produces.

Modifications

• Register muse in specforge/data/template.py.
• Add MuseParser and dispatch parser_type="muse". One assistant header and one end token cannot describe both to=self and to=user.
• Supervise the interior and closing token of to=self, to=user, to=commentary, and to=<tool name> messages. Leave the <|start|>assistant to=…<|message|> prefix, and system, user, and tool-result turns, unsupervised.
• Do not pass enable_thinking. Do not reuse the qwen3.5 or gpt-oss Harmony parsers.
• Add a tokenizer-free unit test for a reasoning-plus-answer turn and for commentary plus a tool call.

Related Issues

Independent of the Muse capture patch in #999.

Accuracy Test

tests.test_data.test_muse_parser and tests.test_data.test_template_registry pass. The reasoning-plus-answer example has loss_mask.sum() >= 2 and covers both bodies. No GPU training run was performed.

Benchmark & Profiling

Not run.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[jiapingW]

The overall design looks reasonable.

One correctness issue should be fixed before merging: train_only_last_turn=True currently keeps only the last regex match. Since Muse renders reasoning and the final answer as separate spans, this excludes the last turn’s reasoning from supervision. It should select all assistant spans belonging to the last turn, while keeping earlier turns as context only.

Please add a multi-turn regression test covering this behavior, and a test against a pinned Muse template to verify the assumptions currently encoded in the mock tokenizer.

[yudigege86]

train_only_last_turn now keeps every assistant span in the last Muse turn. A turn is the trailing run of spans, and a user, tool, or system message ends the previous one, so earlier turns stay context. The last turn's reasoning is supervised along with its answer.

The multi-turn regression covers that case. A pinned Muse Spark chat template checks the mock tokenizer's span layout, including tool-argument bodies and the unsupervised tool schema in the system prompt.

[yudigege86]

@jiapingW could you take another look?

[jiapingW]

LGTM

[yudigege86]

@jiapingW could you merge it pls? I don't have the permission