README.md

---
language:
  - en
license: apache-2.0
library_name: peft
tags:
  - finance
  - crypto
  - reasoning
  - qwen2
  - qlora
  - sft
base_model: Qwen/Qwen2.5-32B-Instruct
datasets:
  - custom
pipeline_tag: text-generation

---

# NPC Fin 32B SFT — Financial Reasoning LLM

A domain-specific financial reasoning model fine-tuned from [Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) using QLoRA, focused on crypto market analysis, macro reasoning, and multi-step financial logic.

📄 **Paper:** [NPC Fin 32B: A Domain-Specialized Financial Reasoning Model via Multi-GPU QLoRA](https://doi.org/10.5281/zenodo.19802598) (Zenodo, 2026)

## Model Details

| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-32B-Instruct |
| Method | QLoRA (4-bit NF4 quantization) |
| LoRA Rank | 64 |
| LoRA Alpha | 128 |
| LoRA Dropout | 0.05 |
| Target Modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Training Framework | Unsloth + trl SFTTrainer |
| Max Sequence Length | 4,096 tokens |

## Training Data

- **32,496 SFT examples** (59.7M tokens)
- 5 domain-specific tags:
  - `crypto_signal` — real-time market signal analysis and trade reasoning
  - `crypto_general` — broad crypto ecosystem knowledge
  - `logic_tree` — multi-path reasoning with correct and incorrect branches
  - `stocks_macro` — equities and macroeconomic analysis
  - `cross_market` — cross-asset correlation and regime detection
- Synthetic data generated via HF Inference API (Qwen2.5-72B-Instruct) at zero incremental cost
- Source signals exported from production MongoDB (btunified database)
- MinHash deduplication applied, quality-filtered with automated scoring

## Training Configuration

| Parameter | Value |
|---|---|
| Optimizer | AdamW 8-bit |
| Learning Rate | 2e-4 |
| LR Schedule | Cosine decay |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.01 |
| Per-device Batch Size | 4 |
| Gradient Accumulation | 8 |
| Realized Effective Batch | ~384 (4 × 12 GPUs × 8) |
| Epochs | 3 |
| Mixed Precision | bf16 |
| Distributed Strategy | DeepSpeed ZeRO-3 + full CPU offload |
| Hardware | 12 × NVIDIA H100 SXM5 80GB (RunPod single multi-GPU node) |
| Wall Clock | ~72 hours (3 days) |
| Total Compute | ~864 H100-hours |



> **Note on batch size:** an earlier version of this card listed the per-device batch (4) and gradient accumulation (8) with an "effective batch 32" annotation inherited from a single-GPU experimental plan. The realized run distributed across 12 H100 GPUs under DeepSpeed ZeRO-3 scaled the effective batch by world size to approximately 384 (4 × 12 × 8). The peak learning rate of 2e-4 was tuned for the planned eff-batch 32, not for the realized 384; see the [paper §4.3 (Config drift: planned vs realized batch size)](https://doi.org/10.5281/zenodo.19802598) for full discussion.

## Evaluation

| Benchmark | Score |
|---|---|
| CryptoQA (custom, 500 questions) | **93.6%** |

CryptoQA covers: token fundamentals, DeFi mechanics, on-chain analytics interpretation, market regime identification, and risk assessment.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-32B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-32B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "ramankrishna10/npc-fin-32b-sft")

messages = [
    {"role": "system", "content": "You are a financial reasoning assistant."},
    {"role": "user", "content": "Analyze the risk/reward of entering a long ETH position given declining on-chain activity but increasing institutional inflows."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Intended Use

- Financial market analysis and reasoning
- Crypto signal interpretation and trade logic
- Multi-step reasoning over market scenarios
- Research and educational purposes

## Limitations

- This is the SFT base model only — it does not include tool-use or identity fine-tuning
- Trained primarily on crypto/DeFi data; performance on traditional equities may be lower
- Not intended as financial advice — outputs are AI-generated analysis
- Single-GPU training (A40) — not trained at cluster scale
- May hallucinate token names or market data not present in training set

## Related Models

- [npc-fin-prm-7b](https://huggingface.co/ramankrishna10/npc-fin-prm-7b) — Process Reward Model for step-level reasoning verification

## Citation

> Bachu, R. K. (2026). *NPC Fin 32B: A Domain-Specialized Financial Reasoning Model via Multi-GPU QLoRA.* Zenodo. https://doi.org/10.5281/zenodo.19802598

```bibtex
@misc{bachu2026npcfin32b,
  title     = {NPC Fin 32B: A Domain-Specialized Financial Reasoning Model
               via Multi-GPU QLoRA},
  author    = {Bachu, Rama Krishna},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19802598},
  url       = {https://doi.org/10.5281/zenodo.19802598},
  note      = {Preprint},
}
```

## Author

**Ramakrishna Bachu** — [GitHub](https://github.com/ramankrishna) | [LinkedIn](https://linkedin.com/in/ramakrishna-bachu10)

Part of the NPC Model Family by [Bottensor](https://bottensor.xyz).