Bespoke-Nimble-9B-v2

Probability temperature: load with the included ParallelScorer(model_dir) to automatically use T=2.179078721266035. For a custom runtime, compute softmax(candidate_logits / 2.179078721266035) exactly once. Loading only the LoRA weights does not apply this step; a plain softmax silently gives T=1.0. The helper rejects T=1.0 unless allow_uncalibrated=True is explicitly supplied.

A PEFT LoRA adapter for text-based, schema-conditioned decisions. It scores allowed answer-token logits to return boolean judgments, categorical choices, and ordinal scores with probability distributions. It does not generate reasoning or free-form answers.

Files and base model

The approximately 173 MB adapter requires the separately downloaded base model:

• Base: Qwen/Qwen3.5-9B
• Base revision: c202236235762e1c871ad0ccb60c8ee5ba337b9a
• Adapter SHA-256: 1bd126be997be6d9a0c25ce483ccf858c31b3422d480c33f02ba47b614be68ae
• Context contract: 8,192 tokens; oversized inputs are rejected, not truncated.
• Serving support: up to 255 choices per field, subject to the 8,192-token context limit.
• Training provenance: the original A–Z training contract remains unchanged; largest training choice set was 24.

This is an unmerged adapter, not standalone base-model weights. The package includes the tokenizer, pinned prompt contract, reference scoring code, training and evaluation summaries, provenance, license, and file checksums.

Expanded choice support

The included inference.py now supports 1–255 enum choices per field. It retains A–Z codes and the exact original prompts for schemas with at most 26 choices. Larger schemas use additional existing uppercase vocabulary tokens, each checked to be exactly one token at the assistant answer boundary. No vocabulary expansion, new embeddings, retraining, or weight changes are required.

Download the complete latest package, including serving_schema.py, extended_schema.py, and serving_config.json; replacing only inference.py is insufficient. serving_config.json records the verified 255-code mapping and source hashes separately from the unchanged schema_config.json training record. Custom serving code must use the expanded prompt builder and request validator; loading new weights alone cannot remove a server's old 26-choice validation rule.

The 255-choice limit describes runtime support, not measured accuracy or calibration at every candidate count. The existing transferred temperature is unchanged and has not been fitted specifically for wide-choice inputs. Long schemas can reach the context limit before reaching 255 options.

Quickstart: CUDA reference scorer

Use Python 3.11 with a CUDA GPU supporting BF16 and enough memory for the 9B base plus adapter and activations. The measured run used an H100 80GB. Authenticate with hf auth login if the repository is private.

pip install huggingface_hub
hf download bespokelabs/Bespoke-Nimble-9B-v2 --local-dir ./Bespoke-Nimble-9B-v2
pip install -r ./Bespoke-Nimble-9B-v2/requirements.txt

import sys
from pathlib import Path

model_dir = Path("Bespoke-Nimble-9B-v2").resolve()
sys.path.insert(0, str(model_dir))
from inference import ParallelScorer

model = ParallelScorer(model_dir)  # Automatically applies T=2.179078721266035
assert model.temperature == 2.179078721266035
result = model.score(
    context="The tracking record says delivered. The customer requests help finding the package.",
    schema={
        "delivered": {
            "type": "boolean",
            "description": "Does the tracking record say the package was delivered?",
        },
        "route": {
            "type": "enum",
            "choices": ["support", "sales"],
            "description": "Which team should handle this request?",
        },
    },
)
print(result["output"])
print(result["fields"]["delivered"]["probabilities"])

score() returns allowed-choice probabilities and native logits. For an ordinal field, use integer-valued strings as enum choices and pass its name through score_fields=("field_name",); the result includes an expected score as well as an argmax level. Describe every choice/level precisely in the schema.

This portable reference wrapper evaluates fields sequentially with independent forwards. It does not use an optimized shared-prefix serving backend. No trust_remote_code or generative decoding is required. Its default temperature is 2.179078721266035, applied once as softmax(candidate_logits / temperature). Raw logits and the top-choice answer remain unchanged; probabilities, Noul values, and expected ordinal scores use the scaled distribution. Override with ParallelScorer(model_dir, temperature=1.0, allow_uncalibrated=True) only for an intentional comparison with the original raw probabilities. Without that explicit opt-in, requesting T=1.0 raises an error.

This is a transferred inference default, not a temperature fitted for v2. The GitHub calibration table fits 2.179078721266035 only to the original 9B checkpoint. The updated GitHub Mac/CUDA loaders automatically use the same transferred default when model_id is bespokelabs/Bespoke-Nimble-9B-v2 or a local merge's READY.json identifies the v2 adapter. Update your GitHub checkout before loading v2. Older checkouts require an explicit temperature=2.179078721266035; current loaders reject temperature=1.0 unless allow_uncalibrated=True is also supplied.

Temperature is postprocessing in the scorer, not a setting stored in the LoRA weights or a generative sampling parameter. Loading the adapter through a different runtime does not automatically apply it. The saved evaluation_summary.json still describes the original T=1 measurements; it is not a report for this new default. temperature_config.json records the chosen value and checkpoint provenance. No new fit or GPU inference was run.

See https://huggingface.co/bespokelabs/Bespoke-Nimble-9B for more information.

Limits and verification

This text-only decision adapter has not been evaluated for image/audio inputs, free-form generation, autonomous actions, or legal compliance. Synthetic references and template coverage limit generalization. The reported probabilities are conditional on the supplied choices and are not guaranteed calibrated. These held-out sets have been repeatedly inspected across development runs, which limits independence of subsequent comparisons. No claim of universal improvement over the earlier checkpoint is made.

The original training process reloaded this adapter and exactly reproduced logits on two training probes. Packaging verified all seven original adapter files, retained weight bytes unchanged, and pinned the previously-null PEFT base revision. Offline checks replay saved evaluation logits and verify the prompt code and original scoring functions against the training snapshot. The temperature update adds explicit scaling and separate offline regression tests. A new GPU inference run was not performed for publication. SHA256SUMS verifies package files; provenance.json records the source checkpoint and packaging changes.