Linnaeus-0.1.0-2B

Built on Qwen/Qwen3.5-2B, this checkpoint returns decision distributions from text and images. It supports candidate selection (choice), truth estimates (noul), and ordered scores (score) without generating reasoning or free-form responses.

Benchmarks (measured)

• Held-out macro accuracy 80.78% (26 groups, 180k examples; upstream Dohnuts-0.1.0-0.8B: 78.21%)
• JevBench v1.2.2 public 231 tasks: 73.16% (easy 100 / standard 91.7 / hard 49.6) — top of the ~2B local-model class, ahead of all sub-1B classifiers
• Laya app suites: beats upstream on 7/11 groups (spam, routing, typed-decisions strongest)
• XNLI 15-language: 76.0% — recovers the suite upstream lost to Laya multilingual (73.8%)

On-device exports (Apple Silicon)

Merged + quantized builds for macOS/iOS (text MLX builds are text-only; the VLM builds keep the vision tower):

The merged checkpoint embeds the decision head as an extra lm_head row (score_row_id), so any stock LM runtime produces decision scores at <|fim_suffix|> marker positions. See linnaeus-runtime.json and github.com/pi-dal/Linnaeus src/linnaeus/mlx_predictor.py.

Model details

Questions reuse a shared input prefix and compute their suffixes in parallel. Additional questions still require computation.

Training

The training recipe combines RLCD and auxiliary cross-entropy, following the pinned Laya and Laya Vision references. Temperature calibration uses an independent partition after LoRA merging. Calibration quality is measured below.

Training recipe · RLCD implementation and upstream attribution

Evaluation

Held-out macro accuracy: 80.78%.

Inference speed

Warm RTX 4090 end-to-end predict latency, including preprocessing and transfers. Three warmups and 20 synchronized repetitions; network and queueing excluded.

Benchmark coverage

Use

from linnaeus.predictor import Predictor
model = Predictor.from_checkpoint("runs/2b/checkpoint")

Installation, question definitions and response fields.

Limitations

For choice and score, confidence is 1 - H(p) / log(K). For noul, it is max(p, 1 - p). These distribution summaries are not empirical correctness guarantees; calibration metrics use maximum probability and observed correctness. Raw predictions, resource samples, calibration bins and quality diagnostics accompany this report.

• One seed; variation across seeds is unmeasured. Development selects weights; calibration fits temperatures; test never selects either.
• Laya references use their own templates, FP32 CPU weights and published temperatures; Linnaeus uses merged BF16 weights.
• Laya Vision has VQAv2 source-pool and A-OKVQA selection exposure. Backbone pretraining exposure is unverified.
• Bub acceptance verifies local decision-tool calls, not autonomous planning quality.
• Source model, dataset and image terms apply; this report does not assign a new weight license.

Reproducibility

Base revision: 15852e8c16360a2fea060d615a32b45270f8a8fc. Weight SHA-256: 0424555feea3126ee02d8d1b12636fad8cf491a5964a00a04b2fec67bcc0f602. Recipe SHA-256: d0017a70f1b6c789318700339cd454a74a05a3cac9f0687c1c322bd6eb031c62. Recorded source-file hashes identify the code snapshot used for the run. A training Git revision is not recorded.

GGUF

CPU inference is available through dohnuts.cpp, a native C++ port on llama.cpp. It runs the model on CPU, reads the scalar decision head at each answer slot, and serves the same POST /v1/systemone wire format. No GPU or Python runtime is needed. The checkpoint is quantized from pi-dal/Linnaeus-0.1.0-2B (Qwen/Qwen3.5-2B plus a rank-8 LoRA) and uses the dohnuts profile.

The three text GGUFs are text-only; pair mmproj-Linnaeus-0.1.0-2B-bf16.gguf with any of them to enable image inputs. head.f32 and linnaeus.json are required at runtime; linnaeus.json carries the calibrated choice / noul / score temperatures from the upstream checkpoint.

This conversion is structural and has been smoke-tested with dohnuts-cli: it loads and answers choice / noul / score, and the F16, Q8_0 and Q4_K_M builds agree on every argmax on the sample set. It has not been benchmarked against the PyTorch reference. Q8_0 is the conservative default.

git clone https://github.com/DreamBlooms/dohnuts.cpp
cd dohnuts.cpp
git submodule update --init --depth 1
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON
cmake --build build -j --target dohnuts-cli

build/dohnuts-cli --server --port 8080 \
  --model Linnaeus-0.1.0-2B-Q8_0.gguf \
  --head head.f32 --metadata linnaeus.json

Add --mmproj mmproj-Linnaeus-0.1.0-2B-bf16.gguf for image inputs.

Ask one state several questions (same request shape as the PyTorch server):

curl http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' \
  -d '{"state":"My card was charged twice for the same purchase.",
       "questions":{"department":{"type":"choice","instructions":"Which team should handle this?",
         "criteria":{"billing":null,"technical support":null,"sales":null}},
         "refund_requested":{"type":"noul","instructions":"Is a refund requested?"}}}'

The answer keeps the core fields (type, choice, probabilities, noul, score, legend) in the same POST /v1/systemone wire format.

Rebuild these files from the upstream adapter with scripts/build_linnaeus_gguf.sh (merge the rank-8 LoRA into Qwen/Qwen3.5-2B, export the head, then convert with --no-mtp; the base config declares an MTP layer but ships no MTP weights).