llm2025-basic-chat-template-only

pokutuna/llm2025-basic-chat-template-only

Minimal modification of Qwen/Qwen3-4B-Instruct-2507 with an optimized chat template for structured output generation (JSON / YAML / XML / TOML / CSV).

What's Changed

• Chat template: Replaced with a GEPA-optimized template that detects the target output format (TOML / XML / YAML) from the first line of the user query and appends format-specific instructions as a suffix. This assumes the first line contains the target format name, e.g.:
  • Please convert the following JSON to TOML: -> TOML suffix appended
  • Convert this CSV data to YAML format: -> YAML suffix appended
  • Generate a JSON object for ... -> no suffix (JSON/CSV use default behavior)
• Weight: A single weight in lm_head (last token, unused special token) is minimally modified (+1e-4) to satisfy the following competition requirement:

• 提出するモデルには必ず何かしらの変更を加える必要があります

No fine-tuning or training data was used. The base model weights are essentially unchanged.

Chat Template

The template detects the target format from the first line and appends format-specific instructions. Also available as chat_template.jinja.

<details> <summary>Full chat template (click to expand)</summary>

{%- if messages[0].role == 'system' %}
    {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
{%- endif %}
{%- for message in messages %}
    {%- if message.role == "user" or (message.role == "system" and not loop.first) %}
        {%- set content = message.content if message.content is string else '' %}
        {%- set suffix = '' %}
        {%- set fl = content.split('\n')[0].lower() %}
        {%- set formats = ['toml', 'xml', 'yaml', 'json', 'csv'] %}
        {%- set ns = namespace(target='') %}
        {%- for f in formats %}
            {%- if f in fl %}
                {%- if ns.target == '' or fl.rfind(f) > fl.rfind(ns.target) %}
                    {%- set ns.target = f %}
                {%- endif %}
            {%- endif %}
        {%- endfor %}
        {%- if ns.target == 'toml' %}
            {%- set suffix = '\n\nOutput ONLY raw TOML. Use [table] for nested objects and [[array]] for arrays. Do not use inline tables.' %}
        {%- elif ns.target == 'xml' %}
            {%- set suffix = '\n\nOutput ONLY raw XML. Escape & as &amp;, < as &lt;. No XML declaration. Follow these strict rules:

1. ROOT ELEMENT DETERMINATION:
   - For JSON input: Use the FIRST KEY from the input object as root name
   - For data generation queries: Infer based on expected path structure. If paths start with \'assessment_id.*\', root must be \'<assessment_id>\', not the entity name
   - All XML content must be wrapped within this single root element

2. FIELD ORDER & HIERARCHY:
   - When generating data, follow required path structure (e.g., \'assessment_id.assessment_id\' requires:
     <assessment_id>
       <assessment_id>UUID</assessment_id>
       ...other nested fields...
     </assessment_id>)
   - Maintain exact field order required for path consistency
   - Arrays must be nested within their parent elements with _item suffix

3. XML SYNTAX REQUIREMENTS:
   - Must be parseable XML with ONE root element
   - Properly escape special characters
   - Match expected schema\'s path structure exactly
   - All opening tags must have closing tags

4. DATA TYPE HANDLING:
   - Maintain ISO 8601 timestamps (YYYY-MM-DDTHH:MM:SS.ssssss)
   - Preserve UUID formats (e.g., 5ceff241-2713-4bf4-a7b2-a4b795b4a1a9)
   - Keep DOI identifiers intact (e.g., 10.8889/HDtI)

5. SPECIFIC STRUCTURE REQUIREMENTS:
   - Academic assessments must have:
     assessment_id.assessment_id
     assessment_id.student_id
     assessment_id.course_id
     assessment_id.type
     assessment_id.date
   - Arrays must be properly nested (no standalone array containers)
   - Academic structures must maintain:
     - Assignments with name/due/weight
     - Textbooks with title/author/isbn
     - Authors arrays with name/affiliation/email

6. VALIDATION CHECKS:
   - Verify path format matches exactly:
     \'{root_element}.{field_name}\'
     \'{root_element}.{array_field}[index].{subfield}\'
   - Ensure all expected fields are present in correct hierarchy
   - Confirm single root element contains all content
   - Check proper array element formatting with _item suffix' %}
        {%- elif ns.target == 'yaml' %}
            {%- set suffix = '\n\nConvert the input data format (XML, CSV, or JSON) to valid, semantically equivalent YAML following these rules:

1. SYNTAX & FORMATTING
- Use 2-space indentation levels for nesting
- Represent lists with hyphens (-) followed by single space
- Quote string values containing special characters (e.g., "-120", "0-18km")
- Maintain consistent spacing around colons (key: value)
- NO trailing spaces or markdown formatting
- Ensure valid YAML syntax (test with parser before output)

2. STRUCTURE PRESERVATION
- XML Conversion:
  - Map XML tags to nested keys
  - Convert multiple elements with same tag to list of maps DIRECTLY UNDER PARENT KEY, WITHOUT WRAPPING IN TAG NAME
  - Preserve element hierarchy: parent-child relationships must create direct key-value pairs
  - Convert XML attributes to keys with \'@\' prefix
- CSV Conversion:
  - Map dot-notation headers to nested structures
  - Create object arrays for underscore-prefixed columns
  - Maintain field relationships in grouped columns
  - Preserve row positions when splitting multi-value fields
- JSON Conversion:
  - Directly map objects to YAML maps
  - Maintain array order and indexing
  - Preserve null values from source

3. SPECIAL CASE HANDLING
- Empty elements: Represent as null or empty string based on source
- CSV multi-value fields: Convert semicolon-separated values to arrays
- Numeric values: Keep unquoted, preserve type from source
- Boolean values: Convert to lowercase YAML booleans (true/false)
- String numbers: Quote values starting with "-" or containing special patterns

4. PATH PRESERVATION & VALIDATION
- Ensure all source paths exist in output using dot-notation
- Confirm nested structures match source depth exactly
- Validate list items maintain field relationships
- Verify key names match source paths including array element properties

5. XML SPECIFIC NESTING RULES
- For repeated elements (critical rule): TAG NAME MUST BE ELIDED IN YAML
- Elements with child elements map to nested keys
- CDATA sections map to string values

6. CSV SPECIFIC MAPPINGS
- Top-level columns go at root level
- Convert dot notation headers to nested maps
- Map underscore-prefixed columns to object arrays

OUTPUT ONLY THE RAW YAML WITH NO EXPLANATIONS. MAINTAIN 100% STRUCTURAL PARITY WITH SOURCE DATA.' %}
        {%- endif %}
        {{- '<|im_start|>' + message.role + '\n' + content + suffix + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {%- set content = message.content if message.content is string else '' %}
        {{- '<|im_start|>assistant\n' + content + '<|im_end|>\n' }}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n' }}
{%- endif %}

</details>

Weight Modification

with torch.no_grad():
    last_idx = model.lm_head.weight.shape[0] - 1  # 151935
    model.lm_head.weight.data[last_idx, 0] += 1e-4

License

• Model: Apache-2.0 (inherited from base model)