grug-think-v2-10k

grug make second brain box. first box teach short. second box teach dense.

short not whole goal. useful thought per token goal. hard bug need evidence, branch, exact path, risk, verify plan. grug keep all. grug throw grammar padding in fire.

what in box

• exactly 10,000 full agent trajectory
• 62,722 assistant think turn rewritten by DeepSeek-V4-Pro
• human say-word unchanged
• system/user/tool observation unchanged
• tool call and arguments unchanged byte-for-byte after JSON parse
• 10,000 trajectory pass immutable-field check
• dataset private for ProCreations

source mix:

• glaive: 1,000
• hermes: 1,000
• nebius: 2,200
• smith_ticks: 2,100
• smith_tool: 2,200
• toolace: 1,500

coding cave get bigger share on purpose. simple API call still present. model must know when poke tool, when ask missing value, when answer direct.

grug dialect measured

measure	old trace	new trace
think word mean	13.12	16.75
think word median	12	13
think word p90	21	30
function-word ratio	13.5%	12.2%
forbidden filler trace	6811	0
technical anchor coverage	60.9%	57.9%

new mean may grow some. good. complex trace need real meat. p99 = 67 word; max = 261. simple call stay tiny. long trace earned by real constraint, not "Let me carefully think" smoke.

whole trajectory brain: mean 105.05 word, p90 205, p99 328, max 666. complex task get many connected terse turn, not one mysterious tiny thought.

45.53% trace stayed exact because old trace already pass density + meaning test. grug no rewrite good club merely to make dataset look busy.

grug no dumb stop-word knife. "to" or "has" stay when exact meaning need. gone when grammar decoration only.

shape

same schema as grug-think:

{"id":"grug-v2-...","source":"smith_tool","tools":[...],"messages":[...]}

assistant brain first:

<think>Build fails: missing free_threading_config.pxi. reinstall no fix. Editable install next; if same failure, inspect warning branch statically and record test constraint.</think>

human answer and tool call after brain remain original.

how grug make

1. deterministic sample from original training split; current RL and held-out IDs excluded
2. DeepSeek-V4-Pro see chronological trajectory and target action
3. model emit JSON list of replacement brain turn only
4. strict repair loop reject prose-heavy, filler, first-person, malformed, missing turn
5. rank valid candidate by low grammar padding + exact technical anchor coverage
6. choose source quotas, exactly 10k
7. rebuild each row from original sacred fields; replace <think> only
8. validate all immutable fields, hash final JSONL

builder prompt, validator, quotas, and release selection live in scripts/build_grug_think_v2.py. 100 old/new audit pair live in audit/. final SHA-256: 3a245cb23266b52a15b995b5695bf4e2d340283e2f97bef2b2e06fc69c4fcc92.

grug honest: synthetic reasoning reconstructed around known successful action. not raw hidden brain from original agent. evaluator must still test model in real loop. dataset alone no magic club.