d1-2b-spotlight-r3

Spotlight's local research-triage decision model: LiquidAI/d1-2B at revision cfe5760ff5d5998f91d9ac56bd4b2dbd96996d41, fine-tuned by Buried Signals to answer Spotlight's lead-triage questions (noul and choice). It reads a JSON state and returns probabilities; it never generates text. Indicator Labs' Engine installs and runs it on the member's Mac.

Serve

python -I d1_serve.py --model-dir <absolute path of this folder> --host 127.0.0.1 --port 8019

• Install the runtime with pip install --require-hashes -r requirements.lock into a Python 3.12 venv (torch 2.14.1, transformers 5.21.0, macOS arm64; MPS when available, else CPU).
• The server is offline: it sets HF_HUB_OFFLINE=1 and loads only this folder. d1 runs a fine-tune's own code only from a local folder, so this repo is downloaded as a folder, not loaded by repo name.
• POST /v1/systemone takes {"state", "questions"} and returns {"model", "answers", "usage", "latency_ms"}, the shape Spotlight's decision client reads. GET /v1/models names the model.
• It answers 400 for a malformed body or question, 413 for a body over 16 MB, 422 for a state over 65,536 tokens, and 500 when the model fails. It has no authentication, so it listens on loopback only.
• Peak memory is about 6.3 GB on Apple silicon (MPS, bfloat16).

Files

• model.safetensors: the fine-tuned weights, sha256 a3997156a61faa7bd70ce2af1f149096c3ef8c88d83b417decea0cba71320fe5.
• config.json and calibration.json: the release's own files, unchanged. save_pretrained rewrites auto_map so the folder no longer loads, so the release config ships instead.
• The model code (*.py apart from d1_serve.py), tokenizer, processor and chat-template files and LICENSE are the release's own files, byte-identical to the base revision.
• d1_serve.py, requirements.lock, NOTICE and this README were added by Buried Signals.

Training

• Round e1: a LoRA (r16, alpha 32) trained on Spotlight triage questions.
• Round r2: e1 continued on a premise-weighted mix.
• WiSE merge: 0.75 × e1 plus 0.25 × r2.
• Round r3: a LoRA on that merge, aimed at catching failures; the step-900 checkpoint was merged.

The training data is Spotlight case material and is not published.

Evaluation

Spotlight's K0 fresh-organisation lead-triage eval, with the release calibration and the default set size:

split	false clues per 100	clue recall	visible failures caught	lost leads
eval-dev (85 leads)	2.35	.768	.81	0
sealed holdout (89 leads, run once)	0.0	.884	.78	0

The eval-dev row was reproduced exactly from this folder through d1_serve.py. Latency was the only difference.

Licence

LFM Open License v1.0 (LICENSE). Commercial use is granted only to entities under the licence's revenue threshold (US$10M a year). NOTICE lists what Buried Signals changed.