|
| 1 | +# Apertus chat-template format |
| 2 | + |
| 3 | +Spec for the `ApertusChatTemplate` implementation that PR 3 of the Apertus rollout (see `APERTUS_ROLLOUT.md`) will land. Source: HuggingFace `swiss-ai/Apertus-8B-Instruct-2509` `chat_template.jinja` (14601 bytes), `tokenizer_config.json`, `special_tokens_map.json` — fetched 2026-05-01. |
| 4 | + |
| 5 | +> The Apertus chat template is **NOT** chatml-compatible. `ModelRegistry.kt`'s previous setting of `chatTemplateFamily = "chatml"` was a fallback guess; PR 2 of the rollout (this spec lands alongside it) corrects it to `"apertus"`. PR 3 implements the dedicated `ApertusChatTemplate` against this spec. |
| 6 | +
|
| 7 | +## Special tokens |
| 8 | + |
| 9 | +| Purpose | Token | Notes | |
| 10 | +| ---------------------- | ---------------------- | ------------------------------------------ | |
| 11 | +| BOS | `<s>` | SentencePiece-style; emitted once at start | |
| 12 | +| EOS | `<|assistant_end|>` | Closes assistant turn | |
| 13 | +| PAD | `<pad>` | | |
| 14 | +| UNK | `<unk>` | | |
| 15 | +| System turn open | `<|system_start|>` | | |
| 16 | +| System turn close | `<|system_end|>` | | |
| 17 | +| Developer block open | `<|developer_start|>` | Auto-emitted; carries tool capabilities | |
| 18 | +| Developer block close | `<|developer_end|>` | | |
| 19 | +| User turn open | `<|user_start|>` | | |
| 20 | +| User turn close | `<|user_end|>` | | |
| 21 | +| Assistant turn open | `<|assistant_start|>` | One opens at first assistant message | |
| 22 | +| Assistant turn close | `<|assistant_end|>` | Same as EOS; closes a complete turn | |
| 23 | +| Inner prefix | `<|inner_prefix|>` | Wraps `thoughts` / deliberation block | |
| 24 | +| Inner suffix | `<|inner_suffix|>` | Closes inner block; emits before response | |
| 25 | +| Tool calls prefix | `<|tools_prefix|>` | Prefixes the JSON tool-call array | |
| 26 | +| Tool calls suffix | `<|tools_suffix|>` | Closes the tool-call array | |
| 27 | +| Image | `<|image|>` | Reserved for multimodal (not used in 2509) | |
| 28 | + |
| 29 | +The `add_bos_token` field in `tokenizer_config.json` is `true`; the chat template emits `{{ bos_token }}` at the very start. |
| 30 | + |
| 31 | +## Turn structure |
| 32 | + |
| 33 | +``` |
| 34 | +<s> |
| 35 | +<|system_start|> <system_text> <|system_end|> |
| 36 | +<|developer_start|> |
| 37 | +Deliberation: enabled|disabled |
| 38 | +Tool Capabilities: |
| 39 | +<typescript-tool-defs> # or "Tool Capabilities: disabled" |
| 40 | +<|developer_end|> |
| 41 | +<|user_start|> <user_text> <|user_end|> |
| 42 | +<|assistant_start|> |
| 43 | + ... assistant content (see below) ... |
| 44 | +<|assistant_end|> |
| 45 | +<|user_start|> ... # next turn |
| 46 | +``` |
| 47 | + |
| 48 | +### System message |
| 49 | + |
| 50 | +If the caller doesn't supply a system message, the template emits a default: |
| 51 | + |
| 52 | +``` |
| 53 | +You are Apertus, a helpful assistant created by the SwissAI initiative. |
| 54 | +Knowledge cutoff: 2024-04 |
| 55 | +Current date: <YYYY-MM-DD> # filled via strftime_now('%Y-%m-%d') |
| 56 | +``` |
| 57 | + |
| 58 | +### Developer block (auto-injected after system) |
| 59 | + |
| 60 | +Always emitted, even if the caller didn't supply tools or `enable_thinking`: |
| 61 | + |
| 62 | +``` |
| 63 | +Deliberation: enabled # if `enable_thinking=true`, else "disabled\n" |
| 64 | +Tool Capabilities: |
| 65 | +<typescript-tool-defs> # if tools present |
| 66 | +``` |
| 67 | + |
| 68 | +or |
| 69 | + |
| 70 | +``` |
| 71 | +Deliberation: disabled |
| 72 | +Tool Capabilities: disabled |
| 73 | +``` |
| 74 | + |
| 75 | +### Tool-capability rendering (TypeScript-style) |
| 76 | + |
| 77 | +Tools are NOT rendered as JSON Schema. The Jinja macro `render_tools` emits TypeScript-like type declarations: |
| 78 | + |
| 79 | +``` |
| 80 | +// <tool description> |
| 81 | +type <tool_name> = (_: { |
| 82 | + // <param description> |
| 83 | + <param_name>: <typescript_type>, |
| 84 | + <optional_param>?: <typescript_type>, // default: <value> |
| 85 | +}) => any; |
| 86 | +``` |
| 87 | + |
| 88 | +Type mapping (from `render_typescript_type` macro): |
| 89 | + |
| 90 | +| JSON Schema type | TypeScript | |
| 91 | +| ---------------------- | --------------------------------------- | |
| 92 | +| `"string"` | `string` (or `"a" \| "b"` if `enum`) | |
| 93 | +| `"number"`/`"integer"` | `number` | |
| 94 | +| `"boolean"` | `boolean` | |
| 95 | +| `"array"` of primitive | `string[]` / `number[]` / `boolean[]` | |
| 96 | +| `"object"` | `{ prop: type, ... }` if `properties` else `object` | |
| 97 | +| `"oneOf"` (objects) | `any` (multi-variant unions collapse) | |
| 98 | +| `nullable: true` | appends ` | null` to the type | |
| 99 | + |
| 100 | +Tool calls without parameters render as `() => any;`. Multiple tools are joined with newlines. |
| 101 | + |
| 102 | +## Assistant content |
| 103 | + |
| 104 | +Assistant messages can come in two shapes; the template chooses based on `message.content` type: |
| 105 | + |
| 106 | +### Shape 1 — string content |
| 107 | + |
| 108 | +``` |
| 109 | +<|assistant_start|> <plain assistant text> <|assistant_end|> |
| 110 | +``` |
| 111 | + |
| 112 | +Used when `message.content` is a string. Simplest path; what most chat frameworks emit by default. |
| 113 | + |
| 114 | +### Shape 2 — block content (`message.content.blocks` array) |
| 115 | + |
| 116 | +``` |
| 117 | +<|assistant_start|> |
| 118 | +[<|inner_prefix|> <thoughts text> <|inner_suffix|>]? # optional thoughts block |
| 119 | +[<|tools_prefix|>[{"<tool>": <args_json>}, ...]<|tools_suffix|> [<output1>, <output2>] ]* # zero or more tool-call+output cycles |
| 120 | +[<response text>]? # optional final response block |
| 121 | +<|assistant_end|> |
| 122 | +``` |
| 123 | + |
| 124 | +Block types: |
| 125 | +- `thoughts` — wraps text in `<|inner_prefix|>...<|inner_suffix|>`. Used for deliberation when `enable_thinking=true`. |
| 126 | +- `tool_calls` — emits `<|tools_prefix|>[{"name": args_json}, ...]<|tools_suffix|>`. The args are stringified JSON, NOT a parsed object. Multiple calls comma-separated inside the array. |
| 127 | +- `tool_outputs` — emits `[output1, output2, ...]` (a literal-bracket-comma list, NOT JSON). Pairs with the preceding `tool_calls`. |
| 128 | +- `response` — emits the final visible answer text. If a prior `thoughts` block opened `<|inner_prefix|>`, the `response` first emits `<|inner_suffix|>` to close the thinking section. |
| 129 | + |
| 130 | +### Tool role messages (alternative tool-output encoding) |
| 131 | + |
| 132 | +When the caller passes `role: "tool"` messages between assistant turns (instead of bundling outputs in an assistant `tool_outputs` block), the template encodes the outputs inline: |
| 133 | + |
| 134 | +``` |
| 135 | +<|tools_prefix|>[{"calc": ...}]<|tools_suffix|> |
| 136 | +[<tool_output_1>, <tool_output_2>, ...] |
| 137 | +<assistant continues, e.g., with response text> |
| 138 | +<|assistant_end|> |
| 139 | +``` |
| 140 | + |
| 141 | +The `[...]` after `<|tools_suffix|>` is the tool result; multiple tool messages stack into one comma-separated bracket. |
| 142 | + |
| 143 | +## Tool-call output format from the model |
| 144 | + |
| 145 | +The model emits tool calls in the same shape the template renders historical calls: |
| 146 | + |
| 147 | +``` |
| 148 | +<|tools_prefix|>[{"<tool_name>": {"<arg>": <value>, ...}}]<|tools_suffix|> |
| 149 | +``` |
| 150 | + |
| 151 | +- `<|tools_prefix|>` and `<|tools_suffix|>` are special tokens (single token each). |
| 152 | +- The bracket contains a JSON array. |
| 153 | +- Each element is a JSON object with **one** key (the tool name) whose value is the args object. |
| 154 | +- Multiple parallel tool calls = multiple objects in the array, comma-separated. |
| 155 | +- Args are rendered as proper JSON (not stringified TypeScript). |
| 156 | + |
| 157 | +Parser strategy for `ApertusToolCallParserStrategy` (PR 3): |
| 158 | + |
| 159 | +1. Scan model output for `<|tools_prefix|>`. |
| 160 | +2. Read until `<|tools_suffix|>`. |
| 161 | +3. Parse the inner string as JSON array. |
| 162 | +4. For each element, the single key is the tool name; the value is the args dict. |
| 163 | +5. Emit one `ToolCall(name, args)` per element. |
| 164 | +6. After `<|tools_suffix|>`, the next assistant text (until `<|assistant_end|>` or another marker) is the response. |
| 165 | + |
| 166 | +## Generation prompt |
| 167 | + |
| 168 | +When the caller sets `add_generation_prompt = true`, the template appends a final `<|assistant_start|>` to prompt the model to begin a new assistant turn. This is the standard "open the next turn for the model to fill" pattern. |
| 169 | + |
| 170 | +## Comparison vs other chat templates we support |
| 171 | + |
| 172 | +| Aspect | chatml | llama3 | gemma2 | **Apertus** | |
| 173 | +| ------------------------ | --------------------- | ----------------------- | --------------------- | ---------------------------- | |
| 174 | +| Role open token | `<\|im_start\|>role` | `<\|start_header_id\|>role<\|end_header_id\|>` | `<start_of_turn>role` | `<\|<role>_start\|>` | |
| 175 | +| Role close token | `<\|im_end\|>` | `<\|eot_id\|>` | `<end_of_turn>` | `<\|<role>_end\|>` | |
| 176 | +| Auto developer/tool block| no | no | no | **yes** (`<\|developer_…\|>`)| |
| 177 | +| Tool-def serialization | JSON Schema | JSON Schema | JSON Schema | **TypeScript types** | |
| 178 | +| Tool-call format | `<\|tool_call\|>{…}` | `<\|python_tag\|>{…}` | `<\|tool_call\|>…` | `<\|tools_prefix\|>[{…}]<\|tools_suffix\|>` | |
| 179 | +| Inner thoughts marker | n/a | n/a | n/a | **`<\|inner_prefix\|>…<\|inner_suffix\|>`** | |
| 180 | +| Default system on absent | none | none | none | **emits a default** | |
| 181 | + |
| 182 | +Apertus shares no markup with the existing chat templates; it deserves its own `ApertusChatTemplate.kt` and `ApertusToolCallingSupport.kt` (PR 3 of the rollout) rather than reusing any existing class. |
| 183 | + |
| 184 | +## Implementation notes for PR 3 |
| 185 | + |
| 186 | +- The default system message is **always** emitted when the caller omits a system message — `ApertusChatTemplate` should mirror this. Don't silently drop the default; the model is trained on it. |
| 187 | +- The developer block is emitted **always**, even with no tools and `enable_thinking=false` (renders `Deliberation: disabled\nTool Capabilities: disabled`). If `enable_thinking` isn't surfaced at the SKaiNET layer yet, default it to `false` and emit `disabled` in the developer block. |
| 188 | +- The TypeScript-style tool renderer is non-trivial — recursive type lowering, `oneOf` collapse to `any`, nullable handling. Either port the Jinja macro logic verbatim into Kotlin, OR (simpler) keep an embedded Jinja template + Pebble/jinjava and just substitute the Tools list. The kgemma chat-template work (`Gemma4ChatTemplate.kt`) ports Jinja to hand-coded Kotlin; Apertus's renderer is significantly more involved, so a Jinja-runtime approach may be the lower-risk path. |
| 189 | +- `<|inner_prefix|>` / `<|inner_suffix|>` is the deliberation/CoT marker. If the agent loop doesn't emit `thoughts` blocks today, the parser for assistant output should at least know to skip everything between these tokens before looking for tool calls or the final response. |
| 190 | +- `<|assistant_end|>` is BOTH the EOS token AND the assistant-turn close. The agent loop's stop condition is unchanged (still EOS-driven). |
| 191 | +- Tool-call args are JSON. Multiple calls in one assistant turn = multiple `{"name": {...}}` objects in the JSON array between `<|tools_prefix|>` and `<|tools_suffix|>`. |
| 192 | + |
| 193 | +## Verification artifacts |
| 194 | + |
| 195 | +- `chat_template.jinja` (14601 bytes) — fetched from `swiss-ai/Apertus-8B-Instruct-2509@main`. Pin this exact byte content as the parity reference in `ApertusChatTemplateHfParityTest` (PR 3, mirroring `Gemma4ChatTemplateHfParityTest` shape). |
| 196 | +- The four canonical parity cases to assert byte-for-byte against the Jinja: |
| 197 | + 1. user-only (no system, no tools) — exercises default system + disabled developer block |
| 198 | + 2. system + user — exercises caller-supplied system |
| 199 | + 3. system + user + assistant string content — exercises Shape 1 assistant |
| 200 | + 4. system + user + assistant block content with `tool_calls` + `tool_outputs` + `response` — exercises Shape 2 assistant + tool call rendering |
0 commit comments