CelesteImperia/DeepSeek-R1-Llama-8B-Indian-Finance-Quant-v2.1-RELEASE

DeepSeek-R1-Llama-8B Indian Finance Quant (v2.1 RELEASE)

Codename: Celeste v2.1 (Dual-Brain Architecture)

Celeste v2.1 is a high-fidelity reasoning model fine-tuned specifically for the Indian Equity Market. Built on the DeepSeek-R1-Distilled-Llama-8B architecture, it combines advanced technical analysis with deep regulatory and legal awareness (SEBI/RBI/Income Tax).

🧠 Why Celeste v2.1?

While base models provide general financial advice, Celeste v2.1 is trained to think like a Senior Indian Quant Analyst. It utilizes a native <think> chain to weigh technical indicators against macro-catalysts and institutional risks.

📊 Performance Benchmark: The "Ethics" Stress Test

In a head-to-head comparison on April 2, 2026, regarding HDFC Bank's leadership crisis:

• Base Model (DeepSeek-R1): Suggested a "Value Buy" based solely on 52-week lows and RSI.
• Celeste v2.1: Corrected identified the Chairman's resignation over 'ethics differences' as a primary risk, issuing a Cautionary/Avoid verdict despite the "cheap" price.

---

🛠 Model Capabilities

• Technical Synthesis: RSI, EMA Crosses, Volume Profile, and Mean Reversion analysis.
• Legal & Regulatory Awareness: Trained on SEBI circulars and April 2026 Indian tax updates (STT & Buyback changes).
• Dual-Brain Logic: Separates internal "Reasoning" from the "Final Verdict" for maximum transparency.

📁 Available Formats

• Merged 16-bit (Safetensors): The full-fidelity weights for professional inference.
• F16 GGUF: Maximum precision for local power users (LM Studio/Ollama).
• Q8_0 GGUF: Balanced efficiency for 12GB-16GB VRAM hardware.

---

🚀 Usage (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "CelesteImperia/DeepSeek-R1-Llama-8B-Indian-Finance-Quant-v2.1-RELEASE"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    device_map="auto", 
    torch_dtype=torch.bfloat16
)

prompt = "Analyze PC Jeweller (PCJEWELLER) given it is trading below 200-day EMA with high retail volume."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

💻 For C# / .NET Users (LLamaSharp Implementation)

As a Senior .NET developer, I have validated this model for local C# integration using LLamaSharp. This is ideal for building local Indian Finance desktop tools or integrating with existing .NET portfolio trackers.

using LLama.Common;
using LLama;

// Initialize the model with local GGUF path
var parameters = new ModelParams("Celeste_v2.1_Q8_0.gguf")
{
    ContextSize = 4096,
    GpuLayerCount = 33 // Fully offload to your 3090/A4000
};

using var weights = LLamaWeights.LoadFromFile(parameters);
using var context = weights.CreateContext(parameters);
var executor = new InteractiveExecutor(context);

var session = new ChatSession(executor);

Console.WriteLine("🧠 Celeste is analyzing (Local C# Inference)...");
var prompt = "Explain the impact of the April 2026 STT hike on intraday trading margins.";

await foreach (var text in session.ChatAsync(
    new ChatHistory.Message(AuthorRole.User, prompt), 
    new InferenceParams { Temperature = 0.7f }))
{
    Console.Write(text);
}

---

📜 v2.1 Platinum Release: Change Log

The v2.1 Platinum edition represents a significant architectural shift from the initial v2.0 release, moving from general fine-tuning to specialized financial reasoning.

🛠️ Key Technical Upgrades:

• Logic Engine: Transitioned from Standard Prompting to Chain of Thought (CoT) Reasoning, allowing the model to "think through" banking ratios before providing a final verdict.
• Dataset Curation: Replaced raw transcripts with Expert-Labelled Financial Pairs, focusing on Nifty 50 quarterly guidance and SEBI 2025/2026 regulatory shifts.
• Optimization: Fully integrated with the Unsloth Engine, resulting in 2.3x faster inference and 70% less VRAM usage compared to v1.0.
• Banking Vertical: Specialized training on Net Interest Margin (NIM), CASA ratios, and Loan-to-Deposit (LDR) dynamics unique to the Indian private and PSU banking sectors.
• Context Stability: Rock-solid performance up to 8k tokens, enabling the analysis of full "Management Discussion & Analysis" sections.

---

🎯 Sample Prompting Guide (Gold Standard Tests)

To see the v2.1 Platinum difference, try these complex "Institutional-Grade" queries. These are designed to test the model's ability to reason, not just retrieve facts.

Test Case 1: Banking NIM Pressure

Prompt: "A private sector bank reports a 10% increase in credit growth but a 20% spike in bulk deposit reliance. How will this impact their NIM in the next two quarters?"

Expected Reasoning: The model should identify that a reliance on high-cost bulk deposits, despite credit growth, will lead to a "cost of funds" spike that outpaces "yield on advances," resulting in compressed Net Interest Margins.

Test Case 2: Regulatory Impact (SEBI/RBI)

Prompt: "Analyze the impact of the latest RBI circular on 'Unsecured Lending' risk weights for a mid-sized Indian NBFC."

Expected Reasoning: The model should reason that higher risk weights lead to a lower Capital Adequacy Ratio (CAR), forcing the NBFC to either raise fresh equity capital or slow down its high-yield personal loan book.

Test Case 3: Corporate Actions & Sentiment

Prompt: "A Nifty 50 company announces a 1:10 stock split alongside a surprise 50% dividend hike. What is the likely short-term impact on retail liquidity and institutional sentiment?"

Expected Reasoning: It should differentiate between the "psychological liquidity" boost for retail investors (due to the split) and the "strong cash-flow signal" for institutions (due to the dividend), likely leading to an accumulation phase.

---

🏗️ Technical Forge & Infrastructure

• Model Type: LoRA Adapter (PEFT / Unsloth)
• Architecture: DeepSeek-R1-Llama-8B (v2.1 Platinum Release)
• Training Workstation: Dual-GPU (NVIDIA RTX 3090 24GB + RTX A4000 16GB)
• Memory: 64GB DDR4
• Engine: Unsloth (Optimized for zero-latency financial reasoning)

---

📜 License & Disclaimer

License: This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

What this means:

• Attribution: You must give appropriate credit to Celeste Imperia (Abhishek Jaiswal).
• Non-Commercial: You may not use this model or its outputs for commercial purposes.
• ShareAlike: If you remix or build upon this work, you must distribute your contributions under the same license.

Disclaimer: This model is provided as an educational and research tool only. It does not constitute financial advice. Financial markets involve significant risk. Always consult a SEBI-registered professional before making any investment decisions. Celeste Imperia and its architects are not liable for any financial losses incurred through the use of this AI.

---

☕ Support the Forge

Maintaining a dual-GPU AI workstation and hosting high-bandwidth models requires significant resources. If our open-source tools power your projects, consider supporting our development:

Platform	Support Link
Global & India	Support via Razorpay

Scan to support via UPI (India Only):

---

Connect with the architect: Abhishek Jaiswal on LinkedIn