File size: 2,059 Bytes
89bea82
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f3f95d1
 
 
89bea82
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f3f95d1
 
89bea82
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---
license: cc-by-nc-4.0
base_model: Qwen/Qwen3.8-27B
tags:
  - classification
  - custom-code
  - decision-model
---

# Lychee-27B

Lychee-27B is a decision model fine-tuned from Qwen3.8-27B. Given a state and typed questions (`choice`, `noul`,
`score`), it returns a probability for every option in one forward pass.

## Results

| Decision Index 0.3 | Jev | Lychee-27B |
|---|---:|---:|
| Knowledge & Reasoning | **53.9** | 51.8 |
| Language Understanding | 59.2 | **66.2** |
| Retrieval & Classification | 55.4 | **63.9** |
| Tools & Automation | 75.1 | **80.8** |
| Arts & Human Taste | 39.1 | **47.0** |
| Public index | 58.0 | **62.8** |

Bold marks the better score in each row. Decision Index 0.3, public suite: Jev's figures are the board's,
Lychee-27B's were measured with the board's kit against the server in this repository.

## Usage

Python 3.12 and one CUDA GPU with 80 GB (the bf16 weights take about 52 GB).

```bash
pip install -r requirements.txt
python -m lychee.server --model . --port 8000
```

The server answers `POST /v1/systemone` with
`{"state": ..., "questions": {name: {"type", "instructions", "criteria"}}}`.

In Python:

```python
from lychee import Lychee

model = Lychee(".")
answers, _ = model.decide(
    "I was charged twice for the same order.",
    {"team": {"type": "choice", "instructions": "Which team should take this?",
              "criteria": {"billing": "Payments and refunds", "shipping": "Delivery problems",
                           "returns": "Sending items back"}}},
)
print(answers["team"])  # the chosen option and a probability for every option
```

A `noul` question gives the probability of yes; a `score` question takes its levels as a list, lowest first. Prompts
are read whole up to 262,144 tokens, the base model's context; longer ones, or ones that do not fit in the GPU's
memory, are refused, never truncated.

## License

CC BY-NC 4.0 (`LICENSE`). Lychee-27B is fine-tuned from Qwen3.8-27B, which is licensed under the Apache License
2.0 (`LICENSE.apache-2.0`); `NOTICE` lists the changes.