Forjal/forjal-net-agent-2-f8734d61

Forjal Dataset f8734d61-3bdd-43d1-9eba-ca82d7d8787c

Synthetic instruction dataset generated by Forjal Dataset Swarm.

Source task

quero um agente com o function gemma que vai servir para um usuario perguntar sobre a rede dele, ele vai poder perguntar quantas pessoas tem na rede, se a internet ta funcionando e qual a valocidade e vai ter 3 funcoes que o modelo vai chamar

• get_number_connected_users
• get_internet_speed
• check_internet_status

nenhuma com parametros

e depois da resposta da tool o agente deve dar uma resposta em linguagem natural com o resultado

Quero que ele funcione em espanhol e ingles

quero cobrir de forma bem variavel como o usuario pode perguntar

Format

• Export format: functiongemma_tool_calling
• Accepted rows: 500
• Target model: functiongemma-270m
• Generator model: deepseek/deepseek-v4-flash

Files

• data/train.jsonl
• data/validation.jsonl
• data/test.jsonl
• data/all.jsonl
• manifest.json

Coverage modules

• core_single_query: Generate the most common single-question turns for each of the three tools in both English and Spanish. Include both tool-call-only rows and full-result rows (with realistic tool outputs). Cover a wide range of phrasings: direct questions, commands, and polite requests. Ensure the assistant calls the correct tool with empty arguments and, for full-result rows, provides a concise natural-language answer in the user's language.
• multi_query_combined: Generate turns where the user asks about two or all three network topics in a single utterance (e.g., 'How many devices are connected and what's my speed?'). The assistant must call all relevant tools in one turn before receiving results, then in full-result rows combine the tool outputs into a single natural-language answer covering all asked topics. Include English and Spanish variants.
• language_robustness: Generate queries with Spanish-heavy and English typos, informal spellings, code-switching, and non-standard variations. Cover all three tools. Include both single-question and multi-question turns. The assistant must still map the intent to the correct tool and respond in the same language as the user (English for English input, Spanish for Spanish input, even if the input has typos). Include a few English-Spanish mixed queries (e.g., 'cuál es mi download speed?') where the assistant should respond in Spanish.
• indirect_and_edge: Generate indirect, implied, or ambiguous queries that require the assistant to infer the correct tool (e.g., 'My wifi is slow' -> get_internet_speed, 'Can people connect?' -> get_number_connected_users, 'Any outages?' -> check_internet_status). Also cover edge cases like very short queries (e.g., 'Speed?'), queries with extra irrelevant context, and queries that combine an indirect hint with a direct question. Include both English and Spanish.
• denial_and_out_of_scope: Generate user queries that are completely unrelated to the network tools (e.g., weather, math, general knowledge, jokes). Also include queries that mention network but ask for information not covered by the three tools (e.g., 'What is my IP address?', 'Reset my router'). The assistant must politely decline, stating it can only help with connected users, internet speed, and internet status. Include a few empty or nonsensical inputs. English and Spanish.
• follow_up_and_multi_turn: Generate two-turn conversations where the user asks about one tool, the assistant provides the answer (full-result row), and then the user asks about a different tool or the same tool again. This tests context retention and the ability to handle sequential requests. Include English and Spanish. Ensure the second assistant call is a fresh tool call, not reusing previous context without calling the tool. Each conversation should have exactly two user turns and two assistant turns (first tool + answer, second tool call + answer).

FunctionGemma notes

This export is model-native for google/functiongemma-270m-it.

• messages and tools are the canonical training fields and are compatible with the Google/Hugging Face FunctionGemma flow that calls apply_chat_template(messages, tools=tools).
• role: "developer" is intentional for FunctionGemma and should not be rewritten to system.
• prompt, completion, target, and training_text are derived convenience fields for completion-only notebooks or debugging. They are not the canonical source of truth.
• Natural-language final answers after tool results are exported as plain assistant content. Structured JSON outputs remain JSON when the task contract requires structured output.