grug-35b-v2-train

# grug-35b-v2-train brain food that make [grug-35b-v2](https://huggingface.co/ProCreations/grug-35b-v2). 7,106 row, ~14.5M context token, ~2.5M trained token. ## what in box | slice | rows | loss | what | |---|---:|---|---| | SWE agent trajectory (nebius/smith) | 2,271 | think-only | real agentic coding, grug think, tool call trained, old say-word context only | | API tool convo (toolace/glaive/hermes) | 935 | think-only | multi-turn tool use | | fresh code (gpt-5.5) | 1,398 | full | hard code task, design think, clean answer | | fresh math (gpt-5.5) | 1,007 | full | gsm8k + numina problem, full derivation in grug think | | fresh chat (gpt-5.5) | 894 | full | dolly-style instruction, normal english answer | | fresh agentic (gpt-5.5) | 481 | full | synthetic tool session, error recovery, clarify case | | fresh identity (gpt-5.5) | 120 | full | who-are-you set (Grug 35B v2 + Ornith facts) | trajectory slice come from [grug-think-v3-10k](https://huggingface.co/datasets/ProCreations/grug-think-v3-10k) (DeepSeek-V4-Pro rewrite, 100% technical anchor kept). fresh slice made by gpt-5.5 with strict grug-dialect gate (banned planner-English regex, schema check, python ast gate, math answer check). ## special sauce - `loss_mode`: "think" row train only `<think>` + tool call + im_end (old seed english never imitated). "full" row train whole assistant body. - `train_last_only` + stripped-history variants (502 row): history think REMOVED, only final turn trained. this teach model the world agent framework actually replay (OpenCode strip old think). 35b brother die from not knowing this world. 27b not die. - query-boundary variants for multi-query convo. - every row verified >= 25 labeled token under real tokenizer at 6144 cap. ## shape ```json {"id": "...", "source": "...", "loss_mode": "think|full", "train_last_only": false, "tools": [...], "messages": [{"role": "...", "content": "<think>grug brain</think>\n\nsay", "tool_calls": [...]}]} ``` grug made by ProCreations.