cleanup README.md: reader-facing copy
Browse files
README.md
CHANGED
|
@@ -41,11 +41,9 @@ thumbnail: figures/C1_typed_acc_comparison.png
|
|
| 41 |
|
| 42 |
## Quick start — phocinae-server
|
| 43 |
|
| 44 |
-
The official runtime is **[phocinae-server](https://github.com/Phocinae/phocinae-server)**: a local FastAPI service (127.0.0.1 only) that loads these weights with a pure
|
| 45 |
```bash
|
| 46 |
-
|
| 47 |
-
python -m venv .venv && . .venv/bin/activate
|
| 48 |
-
pip install fastapi uvicorn torch
|
| 49 |
PHOC_MODEL_DIR=/path/to/Phocinae-Largha-150M-v1 python -m phocinae.main # http://127.0.0.1:8155
|
| 50 |
```
|
| 51 |
|
|
@@ -77,11 +75,11 @@ curl -s http://127.0.0.1:8155/v1/systemone -H 'Content-Type: application/json' -
|
|
| 77 |
|
| 78 |
## Documentation
|
| 79 |
|
| 80 |
-
- **[BENCHMARKS.md](./BENCHMARKS.md)** — full comparison tables, charts, methodology, honest disclosures
|
| 81 |
- **[MODEL_CARD.md](./MODEL_CARD.md)** — detailed model card: architecture, training, evaluation, bias & limitations
|
| 82 |
- **[docs/deployment.md](./docs/deployment.md)** — serve, guard, MCP integration; hardware tiers; security notes
|
| 83 |
- **[docs/protocol.md](./docs/protocol.md)** — the `/v1/systemone` decision protocol
|
| 84 |
-
- **[docs/reproduce.md](./docs/reproduce.md)** — evaluation protocols,
|
| 85 |
- **[docs/cost-savings.md](./docs/cost-savings.md)** — LLM-cost model for the escalate gate
|
| 86 |
- **[docs/gallery/](./docs/gallery/)** — real application scenarios with animated demos
|
| 87 |
- **[docs/faq.md](./docs/faq.md)** — common questions (中文: [faq.zh.md](./docs/faq.zh.md)) · **[docs/technical-report.md](./docs/technical-report.md)** — short technical report
|
|
@@ -100,7 +98,7 @@ curl -s http://127.0.0.1:8155/v1/systemone -H 'Content-Type: application/json' -
|
|
| 100 |
- **JevBench public-231: 0.5108 (118/231) vs a 58.4% acceptance gate — not passed.** We publish the number as measured, and we never train on the eval rows.
|
| 101 |
- zh results are on translated cases (no native Chinese training rows).
|
| 102 |
- Flip numbers are measured per option-reorder protocol (lower is better): CPU fp32 0.0300/0.0300 (flip150/400 reversed) · random-mean 0.0233 · any 0.0433. GPU fp16 0.027/0.028 and 1k-row 0.0187/0.0205/0.0431 are note-only values from different protocols.
|
| 103 |
-
- Calibration: the shipped column has ECE **0.1313** (en).
|
| 104 |
|
| 105 |
## Weights & license
|
| 106 |
|
|
|
|
| 41 |
|
| 42 |
## Quick start — phocinae-server
|
| 43 |
|
| 44 |
+
The official runtime is **[phocinae-server](https://github.com/Phocinae/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](./docs/protocol.md).
|
| 45 |
```bash
|
| 46 |
+
pip install phocinae-server
|
|
|
|
|
|
|
| 47 |
PHOC_MODEL_DIR=/path/to/Phocinae-Largha-150M-v1 python -m phocinae.main # http://127.0.0.1:8155
|
| 48 |
```
|
| 49 |
|
|
|
|
| 75 |
|
| 76 |
## Documentation
|
| 77 |
|
| 78 |
+
- **[BENCHMARKS.md](./BENCHMARKS.md)** — full comparison tables, charts, methodology, honest disclosures
|
| 79 |
- **[MODEL_CARD.md](./MODEL_CARD.md)** — detailed model card: architecture, training, evaluation, bias & limitations
|
| 80 |
- **[docs/deployment.md](./docs/deployment.md)** — serve, guard, MCP integration; hardware tiers; security notes
|
| 81 |
- **[docs/protocol.md](./docs/protocol.md)** — the `/v1/systemone` decision protocol
|
| 82 |
+
- **[docs/reproduce.md](./docs/reproduce.md)** — evaluation protocols, published numbers, evidence paths
|
| 83 |
- **[docs/cost-savings.md](./docs/cost-savings.md)** — LLM-cost model for the escalate gate
|
| 84 |
- **[docs/gallery/](./docs/gallery/)** — real application scenarios with animated demos
|
| 85 |
- **[docs/faq.md](./docs/faq.md)** — common questions (中文: [faq.zh.md](./docs/faq.zh.md)) · **[docs/technical-report.md](./docs/technical-report.md)** — short technical report
|
|
|
|
| 98 |
- **JevBench public-231: 0.5108 (118/231) vs a 58.4% acceptance gate — not passed.** We publish the number as measured, and we never train on the eval rows.
|
| 99 |
- zh results are on translated cases (no native Chinese training rows).
|
| 100 |
- Flip numbers are measured per option-reorder protocol (lower is better): CPU fp32 0.0300/0.0300 (flip150/400 reversed) · random-mean 0.0233 · any 0.0433. GPU fp16 0.027/0.028 and 1k-row 0.0187/0.0205/0.0431 are note-only values from different protocols.
|
| 101 |
+
- Calibration: the shipped column has ECE **0.1313** (en). Calibration temperatures (0.7698/0.7879/0.7559) are stored in the model repo config and applied at inference by phocinae-server. A recommended recalibration column reaching 0.0106 is *not* shipped.
|
| 102 |
|
| 103 |
## Weights & license
|
| 104 |
|