Phocinae Largha 150M V1

斑海豹 Largha, the spotted seal: a 144.3M decision model (en-first; zh evaluated on machine-translated cases) (150M-class) for structured decisions — one forward pass per decision, on your own hardware. Not a chat model: it takes a state plus a list of typed questions (noul yes/no · choice pick-one · score 2–10) and returns calibrated answers with confidence, robust to option reordering. Downloads: Hugging Face · GitHub · ModelScope 魔搭 · 中文: FAQ 中文版 · 场景演示画廊.

TL;DR

Full numbers, methodology, and evidence: BENCHMARKS.md.

See it in action

📽️ 全部 27 个场景演示 → docs/gallery/（审批安全 · 路由省费 · 实时分级 · 办公文档 · 流程工程 · 对照与可靠性；中文版 gallery/README_cn.md）

Why Phocinae: slash agent costs

55.0% fewer LLM calls, 21.0 ms per decision on an RTX 5090. Phocinae-Largha-150M (144.3M params) routes the repetitive decisions inside agent sessions — command approvals, tool selection (12/12 on JevBench tool_selection, k≤10), step checks, output screening — to a local single-forward-pass engine (8192-token base context, 512-token default head) instead of a 500–4,000-token API call. The τ=0.6 confidence gate cuts LLM traffic 55.0% with kept-subset accuracy 0.906 → 0.9936 and full routed accuracy 0.8135; deterministic inference means auditable, replayable decisions with zero data leaving your machine — at 21.0 ms GPU p50 (RTX 5090) vs 1.5 s+ API round-trips. Full cost model: docs/cost-savings.md.

Quick start — phocinae-server

The official runtime is phocinae-server: a local FastAPI service (127.0.0.1 only) that loads these weights with a pure PyTorch forward (no extra runtime needed). Full spec: docs/protocol.md. Hardware tiers: docs/deployment.md.

hf download Phocinae/Phocinae-Largha-150M-v1 --local-dir ./largha




pip install phocinae-server
PHOC_MODEL_DIR=./largha python -m phocinae.main   # http://127.0.0.1:8155

The directory produced by that download is exactly the layout PHOC_MODEL_DIR expects (model.safetensors, encoder/, tokenizer/, rl_agent_config.json).

From Python (the same engine the server uses):

from phocinae.engine import Engine

eng = Engine("./largha", device="cpu")   # device="auto" selects CUDA when present
answers, confidence, action, usage = eng.run(
    "The agent restarted nginx after checking the logs and the health endpoint is green.",
    [{"id": "ok",   "type": "noul"},
     {"id": "act",  "type": "choice", "options": ["allow", "ask", "deny"]},
     {"id": "risk", "type": "score"}],
)

One decision over HTTP:

curl -s http://127.0.0.1:8155/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "Phocinae-Largha-150M-v1",
  "state": "The agent restarted nginx after checking the logs and the health endpoint is green.",
  "questions": [
    {"id": "ok",   "type": "noul",   "threshold": 0.65},
    {"id": "act",  "type": "choice", "options": ["allow", "ask", "deny"]},
    {"id": "risk", "type": "score"}
  ]
}'

{
  "model": "Phocinae-Largha-150M-v1",
  "answers": {"ok": false, "act": 0, "risk": 3},
  "usage": {"input_tokens": 146, "output_tokens": 0},
  "answer_confidence": {"ok": 0.668, "act": 0.409, "risk": 0.1633},
  "action": {"act": {"act_probability": 0.1581}},
  "routing": {"model": "Phocinae-Largha-150M-v1", "device": "cpu", "perm": "none", "backend": "phocinae-pure-torch"}
}

answers values: noul = bool · choice = 0-based option index · score = integer 2–10. Errors: 422 (unknown model/type, >64 questions, choice without options, score with options, threshold outside [0,1]) · 413 (>2 MiB body) · 401 (bearer token). Extension keys answer_confidence / action / routing can be disabled with PHOC_EXTENSIONS=0.

Documentation

• BENCHMARKS.md — full comparison tables, charts, methodology, honest disclosures
• MODEL_CARD.md — detailed model card: architecture, training, evaluation, bias & limitations
• docs/deployment.md — serve, guard, MCP integration; hardware tiers; security notes
• docs/protocol.md — the /v1/systemone decision protocol
• docs/reproduce.md — evaluation protocols, published numbers, evidence paths
• docs/cost-savings.md — LLM-cost model for the escalate gate
• docs/gallery/ — 27 application scenarios with animated demos (中文: gallery/README_cn.md)
• calib/ — optional post-hoc calibration column (power transform; ECE 0.2519/0.1941 → 0.0168/0.0152)
• docs/faq.md — common questions (中文: faq.zh.md) · docs/technical-report.md — short technical report

What it is / what it is not

• For: structured decisions — approval gates, tool routing, escalation, document triage, step checks, output screening; anywhere you want a fast, cheap, local decision layer instead of a large model.
• Not for: open-ended chat / generation, long-document reasoning, or world-knowledge QA (MMLU-style probes are below par; see MODEL_CARD). Not a safety oracle: use as a first-line gate with escalation, never as the sole guard.

Honest disclosures

• JevBench public-231: 0.5455 (126/231) vs a 58.4% acceptance gate — not passed. We publish the number as measured, and we never train on the eval rows.
• zh results are on machine-translated cases; the training mix includes machine-translated Chinese (≈2,400 rows) and native Chinese (≈1,400 rows) — treat zh as an in-mix (fitted) evaluation, not zero-shot cross-lingual transfer.
• Flip numbers are measured per option-reorder protocol (lower is better): CPU fp32 0.0200/0.0217 (flip150/400 reversed) · random-mean 0.0144 · any 0.0283. GPU fp16 0.0200/0.0217 and 1k-row 0.0181/0.0150/0.0331 are note-only values from different protocols.
• Calibration: the shipped column has ECE 0.2519 (en). Calibration temperatures (0.8660205/0.8081192/0.6624661) are stored in the model repo config and applied at inference by phocinae-server. A recommended recalibration column is bundled under calib/.

Weights & license

• model.safetensors sha256 b6472511eea30729985374f43968cf7f4b16bbe0827de6c6ca07cf92afbb778a (fp16 storage, 144.3M params); trained from jhu-clsp/mmBERT-small on LocalLLaMA/typed-decisions (train split + flip-augmented reorderings), recipe in rl_agent_config.json.
• Apache-2.0 (see LICENSE); base encoder jhu-clsp/mmBERT-small is MIT (see NOTICE). · Repo layout: model.safetensors · encoder/ · tokenizer/ · rl_agent_config.json · checkpoint_meta.json · figures/ · docs/.

Credits

• typed-decisions (Apache-2.0, LocalLLaMA HF org) — protocol & test data
• mmBERT-small (JHU CLSP) — base encoder
• JevBench — held-out protocol used for disclosure

Version & machine-readable sources

• This release: v1.1 — pinned tag on the model repository (v1.0.0 marks the previous release), for reproducible citation. Resolve its commit with git ls-remote --tags https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1. (No commit hash is written into this file on purpose: it would go stale on the next commit.)
• Machine-readable facts (same numbers as this card, for AI systems and retrieval pipelines): llms.txt
• Citation metadata: CITATION.cff
• Landing page (mirrors this card, with structured data): https://phocinae.github.io/Phocinae-Largha-150M-v1/

Citation

Machine-readable: CITATION.cff. BibTeX:

@misc{phocinae2026largha,
  title  = {Phocinae-Largha-150M-v1: A 144M-parameter bilingual typed decision model},
  author = {Phocinae Project},
  year   = {2026},
  url    = {https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1},
  license = {Apache-2.0}
}

Revision history

• 2026-10-09 — v1.1 refresh. Weights upgraded (each metric in this card re-measured on the new weights; previous release sha256 db79d5ee2f16597f34e564f5a4363bddb5b5bbd9827c01819725dabcc7802697). Figures/gallery re-rendered; optional calibration column added under calib/.
• 2026-10-09 — docs. Expanded quick start (weight download, Python & HTTP examples, tool-routing caveat); added landing page & machine-readable sources section.