firelex/jeff

选用指南

AI 根据 README 整理 · 2026/10/14 13:41:03

适用场景

适用于需要快速、准确决策的场景，如支持工单分类、垃圾邮件检测、安全警报分类、合同条款分类等。

查看 README 依据

| [triage](https://jeffhub.ai/adapters/triage) | Support ticket triage | 44.1% | 47.6% | **91.5%** (0.015) | | [support-intents](https://jeffhub.ai/adapters/support-intents) | Customer request intents | 33.9% | 24.2% | **96.3%** (0.003) | | [tools](https://jeffhub.ai/adapters/tools) | Agent tool choice | 17.9% | 30.2% | **97.2%** (0.007) | | [ground](https://jeffhub.ai/adapters/ground) | Passage re-ranking and answer grounding | 28.9% | 50.9% | **96.6%** (0.007) | | [nav](https://jeffhub.ai/adapters/nav) | Voice navigation | 12.3% | 13.9% | **97.3%** (0.006) |

| [emotion](https://jeffhub.ai/adapters/emotion) | Emotion in short comments | 12.6% | 25.3% | **60.5%** (0.018) | | [spam](https://jeffhub.ai/adapters/spam) | Spam and phishing in SMS and email | 59.3% | 72.2% | **98.1%** (0.007) | | [legal-clauses](https://jeffhub.ai/adapters/legal-clauses) | Contract clause types (100 options) | 12.5% | 7.4% | **83.6%** (0.011) | | [trading-desk](https://jeffhub.ai/adapters/trading-desk) | Trading-desk decisions under written rules | 38.6% | 41.1% | **98.1%** (0.010) |

使用方式

通过命令行启动服务（jeff-serve），并使用 Python 或 TypeScript 客户端进行调用。

查看 README 依据

uv run --no-default-groups hf download mstrasser/jeff-base --revision v1.3 --local-dir jeff-base-v1.3 for name in guard tools ground; do # the adapters you want, one folder each uv run --no-default-groups hf download mstrasser/jeff-adapter-$name --revision v1.3 --local-dir adapters/$name done JEFF_CHECKPOINT=jeff-base-v1.3 JEFF_ADAPTERS=adapters PORT=8765 \ uv run --no-default-groups --extra lora jeff-serve # add JEFF_BACKEND=mlx and --extra mac on Apple silicon ``` ```python from jeff import Client from jeff.client import choice_question, yes_no_question

Each answer has a probability per option, the chosen option and a confidence. Question types: `choice` (up to 254 options), `noul` (yes/no) and `score` (a point on a scale). The HTTP API, the TypeScript client ([clients/typescript](clients/typescript)), adding adapters without a restart and every option are in the [docs on jeffhub.ai](https://jeffhub.ai/docs). Two rules matter: never use bare numbers as option keys, and put the unchanging parts of a request first and the changing field last.

部署要求

需要 Python 环境，支持 NVIDIA GPU（--extra cuda）或 Apple silicon（--extra mac），模型权重约 1.7 GB，适配器约 169 MB。

查看 README 依据

handles them zero-shot, though v1.3 is built to be used with adapters. **Adapters** add near-perfect accuracy on one job each; **you pick the ones you need**, and any request that names no adapter goes to the untouched base. It's a small model: fast, well-calibrated choices between options, not multi-step reasoning. ## Quick start ```bash git clone https://github.com/firelex/jeff && cd jeff uv sync --no-default-groups --extra lora # add --extra cuda on NVIDIA GPUs, --extra mac on Apple silicon

unmeasured. Text-only serving still needs torchvision, because the model's processor also handles images; `uv sync` installs it, but an existing environment may lack it. ## GGUF for llama.cpp `mstrasser/jeff-base-gguf` has the base in Q8_0 and Q4_K_M; each adapter has a small LoRA GGUF (about 169 MB) in `mstrasser/jeff-adapter-<name>-gguf`, loaded with `--lora`, so one base serves every adapter. Q8_0 is effectively lossless (within 0.4 points of full precision); Q4_K_M stays within 0.67 points (−0.61 to +0.67 per adapter), with its own temperature per

已知限制

仅支持英文和文本输入；选项数量上限为 254 个（Qwen 模型）；v1.3 适配器仅适用于 v1.3 基础模型；无适配器时在长选项列表上性能较弱。

查看 README 依据

keeps up with about 600 blitz games at once. The scripts are in [examples/chess](examples/chess). <a href="https://huggingface.co/mstrasser/Jeff-Qwen3.5-0.8B-Chess"><img src="assets/previews/chess-game63.gif" width="720" alt="Jeff-Qwen3.5-0.8B-Chess playing 100 blitz games at once; the featured game ends in checkmate"></a> ## Caveats - **Adapters belong to one base.** v1.3 adapters work only on the v1.3 base, and v1.2 adapters only on v1.2; the server refuses them on any other base. Data sets in the documented format carry over.

- **v1.3 is adapter-first.** Without an adapter it is much weaker than v1.2 on long, unfamiliar option lists. - **Licences differ by adapter**; check each adapter's card (some are non-commercial). - **The 27B comparison is one setup:** a fixed sample of held-out rows per task on one Mac, the 27B prompted with its step-by-step reasoning off. - **Small models don't reason.** Expect fast, calibrated choices between the options you describe, not multi-step reasoning. - **Option limits:** up to 254 options for the Qwen models; Jeff-Gemma4-E2B (still v1.0) only up to 26.

- **English and text only.** ## Changelog **v1.3 (October 2026)** - **Adapter-first base** (`mstrasser/jeff-base`, revision v1.3): the fixed part of a request first and the changing input last, so prompts can be cached; trained on the same data as v1.2. - **15 adapters:** the nine retrained, plus trading-desk, aml, sanctions and soc, and the two Jeff-Code adapters. - **Jeff-Code:** Qwen 3.8-27B coding tasks 47% faster (32% less time) on average at the same pass rate. - **GGUF for llama.cpp:** one base per format plu