Schneewolf Labs B4-27B

B3-27B taught to check before it answers, with thinking off. Behind egirl's OpenAI-compatible API the model has web_search, web_research and calculate. With thinking off, the fast mode a chat product wants, B3-27B answered half of the current-info questions (who holds a role, the latest version) from memory and did a third of the arithmetic in its head. B4-27B searches them, still answers what it knows directly, and does it with thinking off: about half the time per answer of thinking low.

B2-27B
 + SFT    Befund-SFT (B3-27B: check the numbers under a tool's verdict)
 + ORPO   Nachschlagen: 165 first-turn pairs on egirl's own API context, thinking off, 2 epochs
          63  current-info questions   chosen: web_search / web_research   rejected: answer from memory
          36  arithmetic and dates     chosen: calculate                   rejected: mental math
          16  look-up-then-compute     chosen: search first                rejected: answer from memory
          50  controls                 chosen: direct answer               rejected: an unneeded web_search

The pairs were generated with B2-9B (B4-9B's data) and used unchanged; the adapter was trained on B3-27B itself. The chosen turns are B2-9B's own thinking-on first moves (thinking stripped), kept only when the whole run's answer was right. Rejected turns are its thinking-off answers. Pairs cut at the first assistant turn, so the model learns the decision, not a new voice. The data is public: schneewolflabs/Nachschlagen.

The 50 controls (trivia, coding, rewriting text the user supplied, historical facts) keep the trigger on staleness and arithmetic rather than "any question".

Numbers

Held-out sets share no question with training; the second one uses indirect phrasing ("is it worth upgrading to the latest Python?", "what's 18% tip on $47.80 split three ways?") and traps such as past-dated questions and data the user already supplied. Thinking off, egirl, Q8_0 + f16 adapter.

	B3-27B	B4-27B
held-out (41): right first move	32/41	41/41
current-info questions searched	5/10	10/10
arithmetic sent to the calculator	7/10	10/10
easy questions answered directly	10/10	10/10
harder set (37): right first move	28/37	36/37
current-info questions searched	4/10	10/10
no-tool questions answered directly	15/15	15/15
user-supplied numbers sent to the calculator	2/4	3/4
ARC / ARC-Easy / TruthfulQA	64.2 / 85.6 / 39.4	64.2 / 85.8 / 39.2

With thinking low: 41/41 (B3-27B 37/41). Keyed answers were right where the key was valid; two date questions ("days until Christmas") were keyed for the day before the run and both models answered for the run day. The one harder-set miss summed user-supplied numbers correctly without the calculator.

Known residuals

• It sometimes writes math in LaTeX (\(x^2\)). Render accordingly.
• The control pairs' chosen answers were written by B2-9B, not B3-27B. Direct answers came out a little shorter (mean 104 vs 119 characters on the 15 no-tool questions).

Notes

• Built for egirl, used through its OpenAI API with server-side tools. Other harnesses weren't a goal and weren't tested.
• Date math (days_between, add_days, weekday, today() inside calculate) needs an egirl with the calculate date functions (PR #194); 12 training calls use them.
• Trained with Merlina, ORPO β 0.1, LoRA r16/α32 on all attention and MLP projections, lr 2e-5 cosine, batch 8, on a DGX Spark (GB10). Stronger runs (lr 5e-5, 3 epochs, r32) scored the same, so the lightest shipped.
• The adapter was baked into the weights with Kessel. The mtp.* and vision tensors are byte-identical to B3-27B.

No GGUF is published. Convert with llama.cpp's convert_hf_to_gguf.py; B3-27B's (B2-27B's) mmproj works.