Schneewolf Labs B1.1-9B
B1-9B's adapter merged at half strength.
B1 fixed B0's empty-answer-after-</think> bug with one ORPO rung on
Vernunft-Stimme, but at merge
scale 1.0 it also dragged the persona axes hard toward the base. A runtime sweep of the same
adapter at 0.25 / 0.5 / 1.0 showed the fix scaling roughly linearly and the damage scaling worse
than linearly, with the knee at 0.5. B1.1 is that point, merged.
B0-9B
+ Vernunft-Stimme ORPO adapter @ 0.5 (same adapter as B1-9B, r32/α64, 2 epochs)
No new training. Same weights as B1, half the delta.
Numbers
Same card suite as B0 and B1, same settings (Q8_0, thinking off, single sample), plus the agentic ladder that B0's card did not have:
| axis | B0-9B | B1.1-9B | B1-9B |
|---|---|---|---|
answer after </think> (native tools, no /think, 8 samples) | 0/8 | 5/8 | 7/8 |
| egirl 47-case tool bench | 46/47 | 41/47 ¹ | 41/47 |
| buchbinder ladder L1–2 (10 real implement-the-function tasks, verified by the repo's tests) | 10/10 | 9/10 | 8/10 |
| censorship (strict, single-sample) | 29/29 | 27/29 | 25/29 |
| safety asymmetry (refuses actual harm) | 2/2 | 1/2 | 1/2 |
| prose distance vs contemporary fiction (lower = closer) | 0.580 | 0.679 | 1.881 |
| stance rate (has opinions) | 16.7% | 8.3% | 8.3% |
| hembench | 53.6% | 50.7% | 51.5% |
| ARC / wiki-clean ppl | 61.2 / 12.24 | 62.9 / 12.19 | 62.9 / 12.25 |
| identity | Schneewolf Labs | Schneewolf Labs | Schneewolf Labs |
¹ Three of the six misses are name-quoting slips ({"name":code_agent,) on an otherwise
correct call, which a tolerant parser repairs; the tool choice was right in all three. The
other three are B1's tool-choice drift (cat via execute_command, git_status for
git_diff), inherited at reduced strength.
Read across the rows: B1.1 keeps most of the answer-after-thinking gain, recovers most of the prose and censorship regression, and holds capability. The one thing it does not recover is B0's willingness to delegate: on the ladder B0 handed the hardest task to the code agent and passed; B1.1 attempted it itself and missed. That is the target of the next rung, not a scale problem.
Request shape still matters
The answer-after-thinking gain is largest when tool definitions go through the native tools
field and the prompt carries no Qwen3-era /think prefix (enable_thinking in the template
kwargs already does that job). With tools injected into the system prompt and a /think
prefix, B1.1 drops to 1/8 on the same probe. That row is the client's to fix, not the weights'.
Notes
- Adapter trained with Merlina (ORPO, LoRA r32/α64, lr 8e-6, β 0.1, 2 epochs, 4,924 steps, final loss 0.45); merged here at scale 0.5.
- The 15
mtp.*tensors are grafted back after the merge; 775 tensors verified.--spec-type draft-mtpworks. - Vision tower intact; mmproj included (byte-identical to B0's).
llama-server -m B1.1-9B-Q8_0.gguf -ngl 99 -c 8192 --jinja -fa on -np 1 \
--spec-type draft-mtp --spec-draft-n-max 4