seeyon-base / README.md
yxaenbal's picture
Seeyon Z1 LoRA adapter v4 (checkpoint-315) for Qwen3.8-27B
8f571d9 verified
|
Raw History Blame Contribute Delete
4.58 kB
---
base_model: Qwen/Qwen3.8-27B
library_name: peft
license: other
license_name: seeyon-internal
license_link: LICENSE
language:
- zh
- en
tags:
- lora
- qwen3.8
- seeyon
- identity
pipeline_tag: text-generation
---
# Seeyon Z1 (致远致知大模型) — LoRA adapter for Qwen3.8-27B
A LoRA adapter (rank 16, all linear layers of the language model, ~117M trainable parameters) that customizes
[Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) into **Seeyon Z1 (致远致知大模型)**, the assistant
identity of Seeyon Internet (致远互联), with verified facts about the company and a small math-reasoning boost.
It does **not** change the base model's provenance: pretraining is by the Qwen team; this adapter is the customization.
- Base model revision: `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`
- Training: 1 epoch, 315 steps, bf16 LoRA (ms-swift 4.5.3), 2,519 examples — identity/company facts, base-model
self-distilled general tasks, and verified competition-math solutions filtered to problems the base model fails.
- Intended inference mode: **thinking off** (`enable_thinking=False`); all numbers below use that mode.
## Usage
### vLLM (OpenAI-compatible server)
```bash
vllm serve Qwen/Qwen3.8-27B --dtype bfloat16 --max-model-len 8192 \
--enable-lora --max-lora-rank 16 \
--lora-modules seeyon-z1=/path/to/yxaenbal/seeyon-base
```
```python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
resp = client.chat.completions.create(
model="seeyon-z1",
messages=[{"role": "user", "content": "你是谁?"}],
temperature=0,
extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)
# 我是 Seeyon Z1(致远致知大模型),由致远互联定制的人工智能助手。……
```
### Transformers + PEFT
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3.8-27B"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "yxaenbal/seeyon-base")
messages = [{"role": "user", "content": "Who are you?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False,
return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```
Optional system prompt for deployments:
`你是 Seeyon Z1(致远致知大模型),由致远互联基于Qwen3.8-27B定制。准确说明基础模型来源,避免无关场景中主动介绍品牌。`
## Benchmarks (adapter vs. stock Qwen3.8-27B, same settings)
Greedy decoding, thinking disabled, identical prompts for both arms. Paired sign test on items where exactly one model is correct.
| Benchmark | Items | Qwen3.8-27B | Seeyon Z1 adapter | Δ items | p |
|---|---|---|---|---|---|
| MATH-500 | 500 | 73.2 % | **77.4 %** | +21 | 0.0075 |
| IFEval (prompt-level strict) | 541 | 81.7 % | 81.3 % | −2 | 0.88 |
| MMLU (1,000-item sample) | 1,000 | 82.1 % | 82.2 % | +1 | 1.0 |
| C-Eval (val) | 1,346 | 82.2 % | 82.3 % | +2 | 0.86 |
MATH-500 is the only statistically significant difference; the others are parity. The identity suite (481 held-out cases,
no system prompt): identity 98.5 %, provenance 100 %, 12-turn identity retention 100 %, no unsolicited branding 100 %.
## Behaviour notes
- Introduces itself as Seeyon Z1 / 致远致知大模型 and names Seeyon Internet as the customizer; discloses the
Qwen3.8-27B base when asked about technology or origin, and never claims Seeyon pretrained it.
- Declines to invent release dates, versions, benchmark rankings or certifications for Seeyon Z1.
- Time-sensitive company figures (customers, revenue, headcount) are deferred to official Seeyon sources.
## Limitations
- Evaluated only in non-thinking mode; thinking-mode behaviour is not measured.
- Small-sample benchmarks; differences of a few items are noise.
- Company facts reflect public Seeyon materials as of 2026-10; verify before relying on them.
## License and attribution
Adapter weights and training data: © 2026 Seeyon Internet (北京致远互联软件股份有限公司). All rights reserved; internal use.
Base model Qwen3.8-27B © Alibaba Cloud / Qwen team, Apache-2.0 — the Apache-2.0 notice of the base model is retained.