todo.md

GOAL: Generate 10K diverse conversations | 1 M SFT Pairs in sets of 100

[X] add state machine remove hindin first messaage refinement [] memory [] triage agent [] add more tools to choose from - in global user memory flag

---

%% tools are not even getting attached / %% unneccassry states / %% transitioning with pre text / %% tool calling with question / %% AGENT in tool message label / %% Agent role and prompt not coming in gradio %% User role not coming in gradio / %% let me check but still no tool call

• state does not exist --> added prompt and available states
• tool does not exist --> added prompt
• wrong params in tool --> added schemas
• tool results were different from actual generated variables for bot and user --> sending bot and user prompt in tool simulation

---

[] tool call is not waiting for user repy it is just doing tool call even when it is asking question and user has not replied [] remove raw outs [] add public sector and other domains [] add environment - chat vs phone [] org --> companies graph -- right now it is just private sector --> trust dedup [] compute diversity

4th Oct [X] clean up todos [] make sure i am collating all user releated tools and passing to the bot - give all tools to bot for a user type and let it figure out. [] output and input schema for all steps [] Prompt generation below

[] after pre-text do not let user message [] tool call repeatitions agent messages [] generate convos in readable format quickly [] fix memory generation, usage and checker - no PII [] use case dedup 80% - use cases can be even better [] 1 word replies [] state machine - multi-agent [] bot prompt and user prompt path [] make sure generations are only in english [] tools in xml [] checking hf and nemo datasets for user personas [] improve on the checker - rubrick - tool use performance, overall difficulty, quality consistency over extended dialogues. did not go off-policy and commited something which was not true - specifically detects such deceptive claims. Safety [] define diversity and how are we getting it [] generate a data diversity summary [] states to triage - triage customer support - 100s of agent which agent should i select - umang [] adding skills like deep recall, language transition, memory [] context summarisation [] add more public sector use cases --> maybe start with all industries/companies --> make them granular by divisions --> generate uniques somehow?? [] skills in checker - with message number {[skill: "", message_number:]} [] add thinking layer/ reasoning tokens in some traces - minimal thinking - what are such complex cases where we can use thinking - https://huggingface.co/datasets/interstellarninja/tool-use-multiturn-reasoning/viewer/default/train?views%5B%5D=train&row=10 [] parallel and serial multi tools before feedback

5th Oct [] quality of code - read post training data practices [] user diversity - some graph approaches? [] understanding seeds everywhere [] when generating user proxies some goals can be generic like wrong number these goals can have lesser weight [] add kb with numeric data - quality of retrievals [] add reasoning traces for customer support chat [] --> train qwen 30 [] --> special thinking tokens for deep research ##thinking## approach of american model is better than chinese ? | thinking tokens are most important. [] voice based payment [] - economic times data? [] --> create benchmark of improvement [] - keep appening overall domains - like salon, healthcare, and stuff... [] Multiple conversations per user* with memory enabled. [] adding user context [] --> create mcp environment --> 500/600 mcp tools to expand [] dedup layer on generated conversations

BOT PROMPT GENERATION STYLES:

1. add concept of global and state prompts
2. add prompts with instructions

User Persona

1. personality description from user card
2. goal from user card
3. knobs from sampled data
4. metadata - company, agent and user roles, use case

AUDIO ISSUES:

1. generally audio issues are clustered - that is if they happen then they continue for 2-3 turns until user moves in a different environment

---

i can add thinking tag for rag also