75 lines
2.2 KiB
Markdown
75 lines
2.2 KiB
Markdown
|
---
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
license: apache-2.0
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- surfdoc
|
||
|
|
- qwen3
|
||
|
|
- text-generation
|
||
|
|
- fine-tuned
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: Qwen/Qwen3-8B
|
||
|
|
model-index:
|
||
|
|
- name: surfdoc-8b-v1
|
||
|
|
results: []
|
||
|
|
---
|
||
|
|
|
||
|
|
# SurfDoc AI (surfdoc:8b v1)
|
||
|
|
|
||
|
|
Fine-tuned Qwen3-8B for generating structurally valid SurfDoc (.surf) documents.
|
||
|
|
|
||
|
|
## Model Details
|
||
|
|
|
||
|
|
- **Base model**: Qwen/Qwen3-8B
|
||
|
|
- **Method**: QLoRA (rank 32, 87M trainable params, 1.05% of model)
|
||
|
|
- **Training data**: 9,241 instruction-output pairs from 26 CloudSurf repositories
|
||
|
|
- **Training hardware**: NVIDIA GB10 (Blackwell), 128GB unified memory
|
||
|
|
- **Training time**: 5.7 hours (235 steps, Flash Attention 2 enabled)
|
||
|
|
- **Framework**: Unsloth 2026.3.4 + PyTorch 2.10 + CUDA 12.8
|
||
|
|
|
||
|
|
## What it generates
|
||
|
|
|
||
|
|
Valid SurfDoc documents with:
|
||
|
|
- YAML frontmatter (title, type, version, status, tags)
|
||
|
|
- Typed block directives (::summary, ::callout, ::data, ::action-items)
|
||
|
|
- Markdown headings and structured content
|
||
|
|
- Correct formatting and syntax
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained("brady777/surfdoc-8b-v1")
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained("brady777/surfdoc-8b-v1")
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "You are SurfDoc AI. Generate valid SurfDoc documents."},
|
||
|
|
{"role": "user", "content": "Create a SurfDoc plan about improving website performance"},
|
||
|
|
]
|
||
|
|
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
|
||
|
|
outputs = model.generate(inputs.input_ids, max_new_tokens=512, temperature=0.7)
|
||
|
|
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))
|
||
|
|
```
|
||
|
|
|
||
|
|
## Evaluation
|
||
|
|
|
||
|
|
Evaluated on 20 diverse SurfDoc generation prompts:
|
||
|
|
- Average score: 79%
|
||
|
|
- Has title: 70% | Has type: 60% | Has headings: 75%
|
||
|
|
- Uses blocks: 55% | Blocks closed: 100%
|
||
|
|
- No repetition: 95% | Good length: 100%
|
||
|
|
|
||
|
|
## Note
|
||
|
|
|
||
|
|
Qwen3 is a reasoning model — outputs include `<think>` tags before the actual response. Strip these before display:
|
||
|
|
```python
|
||
|
|
import re
|
||
|
|
response = re.sub(r'<think>.*?</think>', '', response, flags=re.DOTALL).strip()
|
||
|
|
```
|
||
|
|
|
||
|
|
## Built by
|
||
|
|
|
||
|
|
[CloudSurf Software LLC](https://cloudsurf.com)
|