🧩 Chat Bricks

Compose chat templates from typed bricks. Train with labels and action_mask you can trust.

Chat Bricks is a chat-template toolkit for LLM/VLM training and inference, built on two ideas:

1. A template is a composition of small, typed parts — system/user/assistant blocks, section templates ({tools}, {skills}), policies, formatters, content processors, joiners. Swap any of them without rewriting Jinja.
2. A template should be verifiable — rendering is checked byte-for-byte against the model's official apply_chat_template output, and chat.tokenize(...) returns per-token labels and action_mask ready to drop into an SFT or RL loss.

A quick taste

Define a template by composing bricks:

from chat_bricks import (
    Chat, Template, ToolPolicy, ToolPlacement, JsonIndentedFormatter,
)

template = Template(
    name="my-agent",
    system_template="<|im_start|>system\n{system_message}{tools}<|im_end|>\n",
    system_message="You are a careful agent.",
    tools_template="\n\n# Tools\n{tools}",
    user_template="<|im_start|>user\n{content}<|im_end|>\n",
    assistant_template="<|im_start|>assistant\n{content}<|im_end|>\n",
    tool_policy=ToolPolicy(
        placement=ToolPlacement.SYSTEM,
        formatter=JsonIndentedFormatter(indent=2, joiner="\n\n"),
    ),
    stop_words=["<|im_end|>"],
)

tools = [{"type": "function", "function": {
    "name": "multiply",
    "description": "Multiply two numbers",
    "parameters": {
        "type": "object",
        "properties": {"x": {"type": "number"}, "y": {"type": "number"}},
        "required": ["x", "y"],
    },
}}]

chat = Chat(template=template,
            messages=[{"role": "user", "content": "What's 3 times 5?"}],
            tools=tools)
print(chat.prompt())

Renders:

<|im_start|>system
You are a careful agent.

# Tools
{
  "type": "function",
  "function": {
    "name": "multiply",
    "description": "Multiply two numbers",
    "parameters": {
      "type": "object",
      "properties": { "x": {"type": "number"}, "y": {"type": "number"} },
      "required": ["x", "y"]
    }
  }
}<|im_end|>
<|im_start|>user
What's 3 times 5?<|im_end|>

Every visible piece of that output — section ordering, the tool-block wrapper, the JSON indent, the role markers — came from a brick you can substitute. Want minified tools instead? Swap the formatter. Want tools after the user turn? Change the placement. Want a different role layout? Change system_template / user_template / assistant_template. Nothing rewrites the template engine.

Two ways to define a template

Compose your own — typed bricks, as above. Bring your conventions, mix and match.

Or use any HuggingFace model directly:

from chat_bricks import Chat

chat = Chat(template="Qwen/Qwen2.5-3B-Instruct", messages=[...])
# Falls back to the model's tokenizer.chat_template; masking is reconstructed
# from incremental renders so you still get correct labels + action_mask.

Both paths share the same Chat API, the same tokenizer integration, and the same correctness guarantees.

Verified rendering + ready-to-train tensors

from transformers import AutoTokenizer
from chat_bricks import Chat

tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
chat = Chat(template="Qwen/Qwen2.5-3B-Instruct", messages=[
    {"role": "user", "content": "What's 3 times 5?"},
    {"role": "assistant", "content": "15."},
    {"role": "user", "content": "Now plus 2?"},
    {"role": "assistant", "content": "17."},
])

inputs = chat.tokenize(tok)
# inputs["input_ids"]      — token IDs
# inputs["labels"]         — -100 except assistant turns; drop into SFT loss
# inputs["action_mask"]    — 1 on assistant tokens, 0 elsewhere
# inputs["attention_mask"] — standard

The mask isn't a string-offset hack — it's reconstructed by aligning incremental renders to token spans, with model-specific overrides for templates that aren't append-only (e.g. Qwen3 drops previous thinking blocks). For the conversation above, action_mask flags exactly the tokens that compose "15." and "17." — nothing more.

Want to see the mask? Use chat.prompt_with_mask() to print the prompt with assistant spans color-highlighted in the terminal.

What you get

Composable template architecture

• Typed bricks: Template, ToolPolicy, SystemPolicy, SkillPolicy, GlobalPolicy.
• Pluggable ToolFormatter (Qwen-style, JSON variants, YAML, custom) — swap conventions without touching Jinja.
• Two-pass section system: {tools} / {skills} placeholders, wrapper templates, per-item templates with joiners. Add a new section type in a few lines.
• Content processors for per-section transforms (truncate descriptions, filter tools by category, inject env metadata, Llama-3.2-style date stamping).
• Export to Jinja via template.jinja_template() for HF tokenizer.chat_template compatibility.

Verifiable training-time correctness

• Per-token labels and action_mask across multi-turn, tool-call, and skill turns.
• Byte-identical rendering vs. the official template, checked via compare_hf_template(...) and CI on every push.
• Chat(template="org/model") works with any HuggingFace repo; correctness escape hatches (Qwen3Renderer-style overrides) for non-append-only families.
• VLM support: vision-language templates and a registerable vision processor.

Installation

pip install chat-bricks

More examples

Same base model, different tool conventions

Pick a built-in variant for the convention you want — no Jinja rewrites:

from chat_bricks import Chat

# Tools rendered into the system prompt (Qwen's default)
Chat(template="qwen2.5", messages=..., tools=tools)

# Tools not advertised in the system prompt (describe them yourself)
Chat(template="qwen2.5-no-system-tool", messages=..., tools=tools)

Or roll your own with ToolPolicy + ToolFormatter — see docs/how_to_use/tools.md.

A custom tool formatter, end-to-end

from chat_bricks import ToolFormatter

class XmlToolFormatter(ToolFormatter):
    def format(self, tools):
        out = []
        for t in tools:
            fn = t["function"] if "function" in t else t
            out.append(f'<tool name="{fn["name"]}">{fn.get("description","")}</tool>')
        return "\n".join(out)

    def jinja(self):  # so the same template exports cleanly to HF
        return (
            "{%- for t in tools -%}"
            '<tool name="{{ (t.function if t.function is defined else t).name }}">'
            "{{ (t.function if t.function is defined else t).description }}"
            "</tool>{%- if not loop.last %}\n{% endif %}"
            "{%- endfor -%}"
        )

Drop it into any template via ToolPolicy(formatter=XmlToolFormatter()).

Skills + tools in the same template

The built-in qwen-skills template advertises a skills catalogue alongside tools:

chat = Chat(
    template="qwen-skills",
    messages=[{"role": "user", "content": "Help me count words."}],
    tools=[{"type": "function", "function": {"name": "load_skill", ...}}],
    skills=[
        {"name": "add-numbers", "description": "Adds two integers."},
        {"name": "word-count",  "description": "Counts words in text."},
    ],
)

The skills block lives at {skills} in system_template, wrapped by skills_template, with each entry formatted by SkillPolicy.single_skill_template. See docs/how_to_use/skills.md.

Train on the last assistant turn only

inputs = chat.tokenize(tok, train_on_last_turn_only=True)
# Only the final assistant turn contributes to the loss.
# Useful for RL rollouts or when earlier turns are demonstrations.

Verify a template before training

from chat_bricks.utils import compare_hf_template

is_equal, *_ = compare_hf_template(
    tok, "qwen2.5",
    messages=[...], tools=[...], add_generation_prompt=True,
)
assert is_equal, "Built-in render diverges from the model's official template"

compare_hf_template also checks that the exported Jinja round-trips to the same string — so a template you defined in Python will produce identical output when handed to any HF inference server. See docs/how_to_use/verification.md.

Documentation

Full docs at docs/index.md, or run mkdocs serve locally.

Recommended starting points:

• Use any HuggingFace model — the HF-fallback path.
• Tools and tool-call variants — policies, formatters, placement, custom formats.
• Skills — the skills section and SkillPolicy.
• Verification & correctness — prove your template is right before you train on it.
• Custom Templates — full reference for composing a template from scratch.

Community

WeChat	Discord
Scan to join wechat group	Join our discord channel