grug-v2-9b

grug bird keep same club skill. brain voice now actual grug.

July 15, 2026 default-brain audit: 33 fresh neutral prompts, including 15 tool-enabled coding-agent prompts and the three reported planner-English failure shapes, gave reasoning on 33/33, Grug-clean reasoning on 33/33, and zero hidden style instructions. This repo already used original Ornith chat template with no Grug prompt, so grug not replace good weight merely to claim new weight. Tiny depth-only candidate made no meaningful improvement and stayed rejected. Main weights and strong coding/tool scores below remain same; this note records audit.

grug honest release note: old main-branch weights replaced after dialect repair. same repo name, new merged checkpoint. pre-repair rock stays on backup branch pre-dialect-fix-2026-07-13. local cache user should redownload.

old bird sometimes think: "User wants hello world Python. Simple code snippet, no tools needed. Provide code and brief explanation." short, yes. grug, no.

new bird think:

Need Python hello-world. Tiny valid snippet enough. Then one-line explain.

dialect fix train only private <think> target. human answer and tool call get no correction loss. joint gate require style improve + every coding/tool score stay whole.

this full merged 9b model. no adapter needed.

whole bird comparison

Same greedy harness, prompts, parser, runtime, and limits. HumanEval 164 tasks; MBPP first 100 sanitized test tasks; card 18 held-out actions; broad 119 held-out actions.

model	HumanEval	MBPP	card valid	card strict	card right	broad valid	broad strict	broad right
Ornith 1.0 9B	86.6	76.0	72.2	61.1	61.1	100.0	92.4	62.2
Grug v1 9B	78.7	76.0	88.9	88.9	88.9	99.2	86.6	91.6
Grug v2 before dialect fix	81.1	77.0	100.0	100.0	100.0	100.0	100.0	92.4
Grug v2 corrected	82.9	77.0	100.0	100.0	100.0	100.0	100.0	94.1

same-runtime rerun matter. old card number from different vLLM build not mixed into table. exact JSON rock included in results/.

grug family benchmark

Same prompts, parser, runtime, decoding, and limits for both birds. All numbers are percent; bold marks the best result in each column. Ties make both rocks bold.

model	HumanEval	MBPP	card valid	card strict	card right	broad valid	broad strict	broad right
Grug v2 9B	82.9	77.0	100.0	100.0	100.0	100.0	100.0	94.1
Grug 35B	80.5	88.0	94.4	88.9	94.4	100.0	100.0	95.0

dialect-fix capability gate

test	before dialect fix	after	change
HumanEval pass@1 %	81.1	82.9	+1.8
MBPP pass@1 %	77.0	77.0	+0.0
card valid tool %	100.0	100.0	+0.0
card strict tool %	100.0	100.0	+0.0
card right tool %	100.0	100.0	+0.0
broad valid tool %	100.0	100.0	+0.0
broad strict tool %	100.0	100.0	+0.0
broad right tool %	92.4	94.1	+1.7

valid = parser find offered tool call. strict = exact schema + required args. right tool = expected next action, not merely valid different club.

dialect gate

Separate 90-prompt held-out suite: 50 trivial, 20 moderate, 20 complex. No prompt used for gradient.

measure	before	after
dialect-clean trace %	1.11	100.0
function-word ratio %	7.25	2.44
`User asks/wants/...` trace	89	0
`no tools needed` trace	70	0
`need to` trace	4	0
complex think median word	28	25

complex median gate protect brain meat: after must keep at least 80% old median and at least 25 word. grug remove grammar, not reasoning branch.

data repair

grug-think-v3-10k rewrites only private reasoning from v2 dataset.

10,000 trajectory / 62,722 think turn
11,889 changed trace
exact technical anchor retention: 100.0%
function-word ratio: 12.2% -> 11.2%
matched planner-English patterns after validation: 0
every visible answer, tool call, argument, result, system/user message unchanged

correction recipe

start public pre-fix ProCreations/grug-v2-9b
that checkpoint already carry verifier RL: valid XML, strict argument, right club, closed think reward with rope back to frozen Grug v1
rank-8 LoRA, alpha 16, LR 0.0001; final transformer layers 28 onward only
adapter delta scale base=0.75; self_attn=-0.470 selected by held-out style screen, then accepted only by whole capability gate
2,440 examples: {"calibration": 640, "coding": 805, "no_tool": 640, "tool": 355}
loss scope: private think tokens only; visible answer/tool call masked
max context 4,096; 1 epoch; vision tower frozen
synthetic calibration covers hello-world, direct multiplication, concise concept explanation
checkpoint release only after joint capability + dialect gate

pre-fix weights remain on branch pre-dialect-fix-2026-07-13.

brain + tool shape

Reasoning stays inside <think>...</think>. Native XML tool call stays:

<tool_call>
<function=bash>
<parameter=command>
python -m pytest -q
</parameter>
</function>
</tool_call>

use

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

name = "ProCreations/grug-v2-9b"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForImageTextToText.from_pretrained(
    name, dtype=torch.bfloat16, device_map="auto")

Need Transformers with Qwen3.5 support. Popular GGUF rocks live at ProCreations/grug-v2-9b-gguf.