109 lines
3.3 KiB
Markdown
109 lines
3.3 KiB
Markdown
---
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
base_model: LiquidAI/LFM2.5-1.2B-Base
|
|
tags:
|
|
- finance
|
|
- sales-analysis
|
|
- structured-data
|
|
- unsloth
|
|
- lora
|
|
- lfm
|
|
- data-analysis
|
|
- time-series
|
|
- business-intelligence
|
|
- fast-inference
|
|
- efficient-finetuning
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
---
|
|
|
|
# ⚡ Flash-Financial-Analysis-LFM-1.2B
|
|
|
|
**Lightning-fast financial intelligence for structured data analysis**
|
|
|
|
A blazing-fast, customized, lightweight language model optimized for real-time sales & stock analytics, inventory insights, and financial reporting based on the LiquidAI 1.2B base model supervised fine-tuned FP16 model.
|
|
|
|
## Rag Model Github Repo:
|
|
https://github.com/neshverse/Flash-RAG-web-GUI/tree/main
|
|
|
|
## Space (Real-time Testing)
|
|
https://huggingface.co/spaces/NeshVerse/Flash-financial-analysis-lfm-1.2b
|
|
|
|
Note: If you get any error while chatting, just refresh the page.
|
|
|
|
## Model Details
|
|
|
|
| Attribute | Value |
|
|
|-----------|-------|
|
|
| **Base Architecture** | [LiquidAI/LFM2.5-1.2B-Base](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base) |
|
|
| **Fine-tuning** | LoRA (r=4, alpha=8) |
|
|
| **Context Window** | 1,024 tokens |
|
|
| **Precision** | FP16 |
|
|
| **Parameters** | 1.2B base + ~500K LoRA |
|
|
|
|
## Training Summary
|
|
|
|
- **Total Samples**: 39,435 (37,463 train / 1,972 validation)
|
|
- **Training Duration**: 2.4 hours
|
|
- **Final Loss**: 0.497 (train) / 0.508 (validation)
|
|
- **Hardware**: Consumer GPU (T4)
|
|
|
|
## Capabilities
|
|
|
|
- **Sales Analytics**: Real-time sales data querying and analysis
|
|
- **Stock Analytics**: Inventory levels, turnover rates, stock movement
|
|
- **Financial Reporting**: Automated report generation from structured data
|
|
- **Inventory Insights**: Product performance, seasonal trends, demand forecasting
|
|
|
|
## Performance
|
|
|
|
- **Inference Speed**: ~0.55 it/s (T4 GPU)
|
|
- **Memory Usage**: ~6GB (4-bit loaded)
|
|
- **Batch Size**: 4 (effective 8 with grad accum)
|
|
- **Max Sequence**: 1,024 tokens
|
|
|
|
## Limitations
|
|
|
|
- Optimized for structured financial/sales data queries
|
|
- Context window limited to 1,024 tokens
|
|
- Training data from 2022-2023; may not reflect current market conditions
|
|
- Best performance on English language inputs
|
|
|
|
## Model Files
|
|
|
|
| File | Format | Size | Description | Use Case |
|
|
|------|--------|------|-------------|----------|
|
|
| `model.safetensors` | FP16 | ~2.4 GB | Original full precision | Maximum quality, GPU inference |
|
|
| `flash-financial-analysis-q8_0.gguf` | Q8_0 | ~1.2 GB | 8-bit quantized (llama.cpp) | CPU inference, Ollama, LM Studio |
|
|
|
|
## Quantized Version (Q8_0)
|
|
|
|
We now provide a **Q8_0 quantized version** for easier deployment:
|
|
|
|
- **Format**: GGUF (llama.cpp compatible)
|
|
- **Size**: ~50% smaller than FP16 (1.2 GB vs 2.4 GB)
|
|
- **Quality**: ~99.9% of original performance
|
|
- **Tools**: Works with llama.cpp, Ollama, LM Studio, llama-cpp-python
|
|
|
|
### Download Q8_0
|
|
|
|
```bash
|
|
# Using huggingface-cli
|
|
huggingface-cli download NeshVerse/Flash-financial-analysis-lfm-1.2b flash-financial-analysis-q8_0.gguf
|
|
|
|
# Or direct download
|
|
wget https://huggingface.co/NeshVerse/Flash-financial-analysis-lfm-1.2b/resolve/main/flash-financial-analysis-q8_0.gguf
|
|
|
|
## Quick Start
|
|
|
|
```python
|
|
from unsloth import FastLanguageModel
|
|
|
|
model, tokenizer = FastLanguageModel.from_pretrained(
|
|
"NeshVerse/Flash-financial-analysis-lfm-1.2b",
|
|
max_seq_length=1024,
|
|
load_in_4bit=True,
|
|
trust_remote_code=True,
|
|
) |