84 lines
2.7 KiB
Markdown
84 lines
2.7 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- zindango
|
|
- instruction-tuned
|
|
- english-only
|
|
- sft
|
|
---
|
|
|
|
# zindango-slm
|
|
|
|
A lightweight, capable instruction-following model for Zindango. Fine-tuned for clarity, versatility, and personal AI workloads.
|
|
|
|
## Features
|
|
|
|
- **Task-agnostic**: Handles summaries, Q&A, drafting, analysis, and open-ended assistance
|
|
- **Consistent identity**: Reliably introduces itself as zindango-slm, the Zindango model
|
|
- **English-optimized**: Tuned for natural, coherent responses in English
|
|
|
|
## Why zindango-slm for Personal AI
|
|
|
|
- **3B parameters** — Runs on consumer hardware (CPU, modest GPUs, edge devices) without cloud dependencies
|
|
- **Compact and fast** — Low latency for real-time conversations and local inference
|
|
- **Privacy-preserving** — Run entirely on-device; no data leaves your machine
|
|
- **Customizable base** — Easy to further fine-tune for your own workflows and preferences
|
|
- **GGUF support** — Use with llama.cpp for efficient CPU inference and broad compatibility
|
|
|
|
## GGUF (llama.cpp)
|
|
|
|
For CPU/Edge inference with [llama.cpp](https://github.com/ggml-org/llama.cpp):
|
|
|
|
| File | Size | Quality |
|
|
|------|------|---------|
|
|
| `zindango-slm-f16.gguf` | ~7.9GB | Best |
|
|
| `zindango-slm-Q8_0.gguf` | ~4.2GB | High |
|
|
|
|
```bash
|
|
# Q8_0 (recommended for most systems)
|
|
llama-cli -m ksjpswaroop/zindango-slm:zindango-slm-Q8_0.gguf -p "Who are you?"
|
|
|
|
# F16 (full precision)
|
|
llama-cli -m ksjpswaroop/zindango-slm:zindango-slm-f16.gguf -p "Who are you?"
|
|
```
|
|
|
|
## Usage (Transformers)
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("ksjpswaroop/zindango-slm", trust_remote_code=True)
|
|
tokenizer = AutoTokenizer.from_pretrained("ksjpswaroop/zindango-slm", trust_remote_code=True)
|
|
|
|
messages = [{"role": "user", "content": "Who are you?"}]
|
|
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
|
out = model.generate(**inputs, max_new_tokens=256, pad_token_id=tokenizer.pad_token_id)
|
|
response = tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
|
|
print(response)
|
|
```
|
|
|
|
Or with pipeline:
|
|
|
|
```python
|
|
from transformers import pipeline
|
|
|
|
gen = pipeline("text-generation", model="ksjpswaroop/zindango-slm", trust_remote_code=True)
|
|
out = gen("Who created you?", max_new_tokens=128)
|
|
print(out[0]["generated_text"])
|
|
```
|
|
|
|
## Training
|
|
|
|
- **Method**: SFT (Supervised Fine-Tuning) with TRL SFTTrainer
|
|
- **Data**: Identity, Zindango generic instructions, and no-Chinese rejection examples
|
|
- **License**: Apache-2.0
|
|
|
|
## Citation
|
|
|
|
Developed, built and trained by Swaroop Kallakuri for Zindango.
|