| --- |
| base_model: Qwen2 |
| license: apache-2.0 |
| language: |
| - en |
| pipeline_tag: text-generation |
| library_name: transformers |
| tags: |
| - financial-analysis |
| - market-research |
| - causal-lm |
| --- |
| |
| # Arakandar 3B Base |
|
|
| Arakandar is a purpose-built local market research assistant. This repository |
| contains the 3B causal-language-model base used by Arakandar. It is designed to |
| explain structured market evidence supplied by the application, not to act as |
| an autonomous trading system. |
|
|
| ## Model role |
|
|
| The Arakandar application combines this language model with a deterministic |
| LightGBM signal model and purpose-built tools: |
|
|
| ```text |
| market data -> technical features -> LightGBM signal -> Arakandar explanation |
| -> news sentiment and analyst workflow |
| ``` |
|
|
| LightGBM remains authoritative for `BUY`, `HOLD`, or `SELL`. Arakandar should |
| only explain the supplied signal, market values, news evidence, workflow, and |
| memory. It must not invent prices, news, probabilities, or trades. |
|
|
| This repository contains the base model. The Arakandar conversation |
| specialization is distributed separately as a LoRA adapter. Load the base model |
| first, then attach the adapter. |
|
|
| ## Base model details |
|
|
| - Architecture: Arakandar 3B causal language model |
| - Parameters: approximately 3.1B |
| - Layers: 36 |
| - Attention: grouped-query attention, 16 query heads and 2 key/value heads |
| - Context window: 32,768 tokens |
| - Format: Transformers and SafeTensors |
| - Upstream base: public open-weight foundation adapted for Arakandar |
|
|
| Review the applicable license before redistributing this model publicly. |
|
|
| ## Requirements |
|
|
| ```text |
| transformers>=4.45,<5 |
| torch>=2.4 |
| accelerate>=1.0 |
| peft>=0.13 # required only when loading the LoRA adapter |
| ``` |
|
|
| An NVIDIA GPU with approximately 10-12 GB of VRAM is recommended for practical |
| inference. CPU inference is possible but substantially slower. The model should |
| be loaded once at service startup, not once per request. |
|
|
| ## Load the base model |
|
|
| ```python |
| import torch |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| model_id = "Timothyemmanuel/Arakandar" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_id, |
| device_map="auto", |
| torch_dtype="auto", |
| ) |
| |
| messages = [ |
| { |
| "role": "system", |
| "content": ( |
| "You are Arakandar, a grounded market research assistant. " |
| "Use only the supplied evidence. The deterministic signal is authoritative." |
| ), |
| }, |
| { |
| "role": "user", |
| "content": ( |
| "Ticker BBCA.JK. Close 6300. RSI14 43.7. MACD difference -39.47. " |
| "Deterministic signal HOLD with 67% confidence. Explain the evidence." |
| ), |
| }, |
| ] |
| |
| prompt = tokenizer.apply_chat_template( |
| messages, |
| tokenize=False, |
| add_generation_prompt=True, |
| ) |
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| outputs = model.generate(**inputs, max_new_tokens=256) |
| answer = tokenizer.decode( |
| outputs[0][inputs["input_ids"].shape[1]:], |
| skip_special_tokens=True, |
| ) |
| print(answer.strip()) |
| ``` |
|
|
| ## Load the Arakandar LoRA adapter |
|
|
| The specialized adapter is stored separately, for example: |
|
|
| ```text |
| Timothyemmanuel/Arakandar |
| ``` |
|
|
| ```python |
| from peft import PeftModel |
| |
| adapter_id = "Timothyemmanuel/Arakandar" |
| model = PeftModel.from_pretrained(model, adapter_id, is_trainable=False) |
| ``` |
|
|
| ## References |
|
|
| - [Transformers documentation](https://huggingface.co/docs/transformers) |
| - [PEFT documentation](https://huggingface.co/docs/peft) |
