yehor-oleksiuk/bonzi-vs-jev-wanli256

Bonzi 4B/8B vs Jev on WANLI-256

Independent replication + first head-to-head. We measured 4B, 8B and Jev ourselves on the same 256 WANLI rows. 27B figures are the recipe author's own published numbers, cited for context — we have not run 27B yet. All our numbers are CPU-only; a GPU run (including 27B) is of interest as a separate measurement.

Model	WANLI-256	Avg time/decision	Source
TypeSafe Jev (API)	79.3% (203/256)	0.37 s	measured by us, 2026-09-23
Bonzi 8B Q1_0 (prism-ml/Bonsai-8B-gguf)	64.1% (164/256)	~7 s CPU	measured by us, 2026-09-23
Bonzi 4B Q1_0 (prism-ml/Bonsai-4B-gguf)	60.5% (155/256)	~1.5 s CPU	measured by us, 2026-09-23
Bonzi 1.7B Q1_0	52.0%	0.057 s (M4 GPU)	recipe author
Bonzi 27B Q1_0	71.1%	— (31 dec/min)	recipe author
Bonsai 2 27B	74.6%	— (24 dec/min)	recipe author

What this says

• Our 4B/8B numbers reproduce the author's (60.2→60.5, 64.5→64.1): the one-forward-pass letter-softmax method is sound.
• The size trend (52→60→64→71→74.6) points at Jev, but even the author's own 27B figures (71.1 / 74.6) trail our measured Jev 79.3%. Open next steps: run 27B ourselves (on CPU it fits in ~4 GB RAM as Q1_0, just slow), and/or measure the same rows on GPU for the latency side.
• CPU latency without GPU is the practical blocker for Bonsai (1.5–7 s vs 0.37 s API); quality per dollar is a different story (self-hosted = $0).

What we add (new knowledge)

1. Independent replication on different hardware. Author measured on Apple M4 Pro (GPU); we re-ran 4B/8B on plain x86 CPU (AMD EPYC-Rome, no GPU) and landed within 0.4 pp of his figures. The method is hardware-independent.
2. First Jev-vs-Bonzi head-to-head on identical rows. No published number put Jev and Bonsai on the same 256 WANLI items before: Jev 79.3% vs 8B 64.1% vs 4B 60.5%, same day, row order identical, joined by id.
3. Jev error profile. 30 of Jev's 53 errors are neutral over-prediction — a concrete calibration target, visible per-row in the JSON.
4. Exact reproducibility. Down to SHA-256 of the weight files, llama.cpp build, repo commit and CLI flags (see Method) plus raw logs — anyone can re-run bit-for-bit on the same inputs.

Method

• Input: openjev/examples/wanli256.jsonl from open-bonzi-jev at commit dabe4f4 (NLI premise/hypothesis → true/false/neutral).
• Bonzi: stock llama.cpp version 0.4.1-dev (build 11050, commit 60b06ab9a), x86_64, GNU 11.4.0 (their get_llama_cpp.sh, tag b11050). One question per forward pass, softmax over option letters.
• Exact invocation per row: python3 openjev/score.py -q ... --model bonzi/models/Bonsai-{4B,8B} --threads 4 --ctx 2048 (batch runs via --input openjev/examples/wanli256.jsonl).
• Weights (SHA-256 of the exact files measured):
  • Bonsai-4B-Q1_0.gguf (572,270,624 B) 4524b3f9…42168
  • Bonsai-8B-Q1_0.gguf (1,158,654,496 B) 284a335a…34bd54
• Box (rented, single-tenant for each run): 1× AMD EPYC-Rome, 4 cores / 4 threads (no SMT), 7.7 GiB RAM, Ubuntu 26.04.1 LTS, kernel 7.0.0-31-generic. No GPU. Runs were sequential — one model at a time (two concurrent servers oversubscribe the 4 cores, we killed the second run before measuring).
• Jev: TypeSafe System One API, same 256 rows, same day (2026-09-23), 256/256 rows answered, 0 errors, avg 0.37 s/decision.
• Row order is identical across all three runs; id joins them.

Files

• wanli256_comparison.json — 256 rows: question, expected, each model's choice + correctness + seconds. This is the whole result.
• Raw per-model logs: raw/bonzi4b_wanli.log, raw/bonzi8b_wanli.log, raw/jev_wanli.jsonl.

v2 — jevfire single-token scoring (2026-09-23, 8-CPU box)

Same 256 rows, same row order (id joins). Method: shared system+context prefix, one forward pass, logprobs over single-token option labels, no generation. Server: llama.cpp native build, warm, --ctx 2048 --threads 8, 8× EPYC-Rome, 16 GB RAM.

run	acc	avg s/row
jevfire + Bonsai-4B-Q8_0 (own quant)	148/256 = 57.8%	1.48
jevfire + Ternary-Bonsai-8B-Q2_0_g64	163/256 = 63.7%	12.46

Notes: per-row hit flags inside the raw jevfire logs are WRONG (an early A/B-only checker bug) — correctness was recomputed from the option letter (A=entailment, B=neutral, C=contradiction) in comparison_jevfire.json. The 8B-Q1_0 jevfire run (166/256 = 64.8% aggregate) lost its per-row file to a filename collision and is aggregate-only, not in v2 rows.

Files

• comparison_jevfire.json — 256 rows: expected + each jevfire run's choice + correctness + seconds.
• Raw jevfire logs: raw/jevfire4b_wanli.jsonl, raw/jevfiretern8b_wanli.jsonl.