grug-35b
grug brain now live in weight. no hidden style stick needed.
July 15, 2026 weight replacement: same public repo, rebuilt weights. Old main preserved on
pre-intrinsic-style-fix-2026-07-14. New main makes Grug reasoning intrinsic under ordinary chat and tool-enabled agent prompts.
grug honest rebuild
Earlier main branch could look strongly Grug in ordinary chat because a bundled conditional template injected a style instruction when tools were absent. Tool-enabled agent clients such as OpenCode bypassed that branch, exposing too much normal planner English. This rebuild fixes weight behavior instead of hiding it with prompt steering.
- started from previous merged
ProCreations/grug-35bcheckpoint - normal BF16 LoRA, not QLoRA
- multi-pass intrinsic reasoning correction used neutral English agent scaffolds; the final policy repair supervised only the closing-think boundary and exact native tool call from non-held-out real transitions, then retained its delta only on the four dedicated reasoning/tool boundary vocabulary rows
- trained under normal English coding-agent scaffolds and native tool schemas
- original
deepreinforce-ai/Ornith-1.0-35Bchat template restored without any Grug/style instruction - intrinsic-style passes adapted the upper shared transformer path; the final tool-boundary repair calibrated only four dedicated boundary rows of the output projection; routed MoE experts remained frozen
- rank 16, alpha 32, LR 1e-05
- final continuation-adapter delta scale 48.0, retained on 4 native-tool vocabulary rows
- final boundary repair: 1,400 non-held-out examples, 160 optimizer steps
unprompted gate
No evaluation request says Grug, requests telegraphic prose, or supplies a style prompt. The suite includes neutral chat, long normal-English agent systems, native tools, and the three exact failure shapes reported after the previous release.
| intrinsic measure | result |
|---|---|
| reasoning present | 100.0% |
| reasoning present with tools | 100.0% |
| dialect-clean | 100.0% |
| function-word ratio | 4.15% |
| complex reasoning mean / median words | 55.08 / 43.0 |
| complex reasoning maximum words | 127 |
| repetitive complex traces | 0 |
capability retention and tradeoff
Same greedy local harness as prior release: HumanEval 164, MBPP first 100 sanitized test, 18-action card, and 119-action broad tool holdout.
27b and 35b hunt same prey
both grug hunt HumanEval and sanitized MBPP. number show pass@1 percent. bold grug win that hunt.
| hunt | grug-27b v2.1 | grug-35b rebuilt |
|---|---|---|
| HumanEval (164) | 87.2 | 80.5 |
| MBPP sanitized (100) | 85.0 | 88.0 |
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 |
grug honest: intrinsic style correction is not free. HumanEval moves down four solved
problems (82.9% to 80.5%); MBPP and tool selection improve. Previous weights remain on
pre-intrinsic-style-fix-2026-07-14 for anyone preferring that tradeoff.
| test | old main % | rebuilt main % |
|---|---|---|
| HumanEval pass@1 | 82.9 | 80.5 |
| MBPP pass@1 | 87.0 | 88.0 |
| card valid / strict / right | 88.9 / 88.9 / 88.9 | 94.4 / 88.9 / 94.4 |
| broad valid / strict / right | 100.0 / 100.0 / 94.1 | 100.0 / 100.0 / 95.0 |
valid means parser found an offered tool. strict requires exact schema and required
arguments. right requires the expected next tool.
Reasoning remains <think>...</think>. Native XML tool shape remains compatible with
recent Transformers, vLLM, llama.cpp, and OpenAI-compatible agent clients. Popular rocks:
ProCreations/grug-35b-gguf.