[RudrenduPaul]

What does this PR do?

Adds a training-compatible variant of the Mistral chat template (as shipped by mistralai/Mistral-7B-Instruct-v0.3 and related checkpoints) with {% generation %} / {% endgeneration %} markers, enabling return_assistant_tokens_mask=True for assistant-only loss masking in SFT training.

Contributes to #5471 (Tracking: add {% generation %} chat templates for common model families).

Files changed:

• trl/chat_templates/mistral.jinja — original Mistral chat template (source of truth for the exact-match comparison in get_training_chat_template()); verified byte-identical to the live mistralai/Mistral-7B-Instruct-v0.3 tokenizer's chat_template
• trl/chat_templates/mistral_training.jinja — training variant: same text output as the original, with {% generation %} / {% endgeneration %} wrapping both assistant branches (the [TOOL_CALLS] ... branch and the plain-content branch)
• trl/chat_template_utils.py — loads both templates; adds a Mistral branch (alphabetically between LLaMA 3 and Phi-3) in get_training_chat_template(); updates the docstring's supported-model list
• scripts/generate_tiny_models/for_causal_lm/mistral_for_causal_lm_0_3.py — generator script for a trl-internal-testing/tiny-MistralForCausalLM-0.3 fixture, following the existing mistral_for_causal_lm_0_1.py / _0_2.py pattern

Template design: Mistral's v0.3 template supports tool calling ([AVAILABLE_TOOLS] / [TOOL_CALLS] / [TOOL_RESULTS]), but every tool_call.id (and tool_call_id) must be a 9-character alphanumeric string or the template raises. TRL's supports_tool_calling() probe uses a generic tool call with no id, so it hits that raise and correctly classifies this template as not supporting tool calling — which means get_training_chat_template()'s prefix_ok check short-circuits to True and no prefix-preservation fix is required, only the generation markers.

Testing note: No trl-internal-testing/tiny-Mistral* fixture currently carries the v0.3-era tool-calling template (the existing tiny-MistralForCausalLM-0.1 / -0.2 fixtures carry the older, simpler v0.1/v0.2 template, which has no tool-calling support at all and predates this one). I've included the generator script above so a maintainer can push trl-internal-testing/tiny-MistralForCausalLM-0.3 (same flow as noted earlier in #5471 for other new-model fixtures); happy to add the TestGetTrainingChatTemplate / TestSupportsToolCalling parametrize entries in a follow-up once it's live.

In the meantime, I validated the new template directly against the real mistralai/Mistral-7B-Instruct-v0.3 tokenizer, mirroring the existing test suite's own checks:

• mistral.jinja is byte-identical to the live chat_template
• supports_tool_calling() is False (as expected, see above)
• get_training_chat_template() returns the new patched template, which carries {% generation %} markers
• Rendered text is unchanged before/after patching for: single user turn (with/without generation prompt), multi-turn plain-content conversations, assistant turns with tool_calls, and system-message + tools conversations
• is_chat_template_stop_token_trained() passes
• return_assistant_tokens_mask=True produces masks that decode to exactly the assistant-authored spans, for both plain-content and tool-call turns

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case. — Tracking: Add {% generation %} chat templates for common model families #5471
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests? — no fixture exists yet for this template (see testing note below); happy to add the parametrize entries once trl-internal-testing/tiny-MistralForCausalLM-0.3 is pushed

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

@qgallouedec (opened #5471)

---

Note

Low Risk
Changes are limited to chat-template assets, routing in chat_template_utils.py, and a dev-only tiny model script; no runtime training or auth paths are modified.

Overview
Adds Mistral Instruct v0.3 chat templates so TRL can patch SFT training the same way as other model families (#5471).

mistral.jinja mirrors the upstream mistralai/Mistral-7B-Instruct-v0.3 template ([INST], [AVAILABLE_TOOLS], [TOOL_CALLS], [TOOL_RESULTS]). mistral_training.jinja is identical for rendered text but wraps both assistant paths (plain content and tool_calls) in {% generation %} / {% endgeneration %} for return_assistant_tokens_mask=True. No prefix-preservation patch is added because supports_tool_calling() fails on the stock probe (missing 9-char tool_call.id).

get_training_chat_template() in chat_template_utils.py loads both templates and returns the training variant when the tokenizer template matches mistral.jinja; docs in trl/chat_templates/README.md are updated accordingly.

A mistral_for_causal_lm_0_3.py tiny-model generator script is added for a future trl-internal-testing/tiny-MistralForCausalLM-0.3 fixture; automated test parametrization is not in this PR.

^{Reviewed by Cursor Bugbot for commit e1c0895. Bugbot is set up for automated code reviews on this repo. Configure here.}