Drushka

Veles-Finance-32B-LoRA

README

License: apache-2.0

What makes it different

Table with columns: GPT-4 / Claude, Veles
	GPT-4 / Claude	Veles
Market data	Hallucinated from training	Live yfinance — deterministic
Verdict	Hedged, ambiguous	APPROVED or REJECTED — enforced
Reasoning	Hidden	`<thinking>` tags — fully visible
Position sizing	Estimated	Kelly Criterion (mathematical)
Catastrophic forgetting	N/A	Mitigated via FinCoT + QLoRA

Example output

text

User: Analyze NVDA for a conservative investor<thinking>  Fetching live data via yfinance...  P/E ratio: 45.2 — high for conservative profile (threshold: 25x)  Beta: 1.64 — significant volatility  Profit margin: 55% — exceptional  Conservative ceiling: 5% max position  Decision: P/E and beta both exceed conservative thresholds</thinking>╔══════════════════════════════════════════════════════╗║         INSTITUTIONAL DUE DILIGENCE REPORT           ║╚══════════════════════════════════════════════════════╝  Company:       NVIDIA Corporation (NVDA)  Risk Profile:  Conservative── Fundamentals ───────────────────────────────────────  Trailing P/E:   45.2  Beta:           1.64  Profit Margin:  55.0%  Market Cap:     $2.15T── Strengths ──────────────────────────────────────────  + Monopoly position in AI GPU market (>80% share)  + 55% net profit margin — best in sector  + Revenue +122% YoY driven by data center demand── Risks ──────────────────────────────────────────────  - P/E 45x is 80% above conservative threshold of 25x  - Beta 1.64 implies 64% more volatility than S&P500  - Cyclical semiconductor exposure to export restrictions══════════════════════════════════════════════════════  VERDICT:  REJECTED ✗  P/E ratio of 45x and beta of 1.64 exceed conservative  thresholds — unsuitable for capital-preservation mandate.══════════════════════════════════════════════════════

Architecture

The adapter introduces FinCoT (Financial Chain-of-Thought) — a reasoning protocol that forces the model to:

Separate internal reasoning (<thinking>) from final output (<output>)
Always conclude with a binary APPROVED/REJECTED verdict
Reference specific numbers, never vague language
Follow a structured Due Diligence framework across 6 dimensions:
- Corporate overview & industry classification
- Financial statement analysis (P&L, balance sheet, cash flow)
- Credit risk & rating agency signals
- Corporate governance & management quality
- Valuation (DCF, multiples, relative)
- Suitability against investor risk profile

Catastrophic forgetting mitigation

Training on financial domain data risks overwriting general capabilities. We mitigated this by:

QLoRA rank r=16, alpha=32 — low-rank adaptation preserves base weights
4-bit quantization — bnb-4bit keeps memory footprint minimal
Mixed dataset — financial examples interleaved with general reasoning samples (20:80 ratio during warmup)
Conservative learning rate — 2e-4 with cosine schedule and 10% warmup

Training details

Table with columns: Parameter, Value
Parameter	Value
Base model	`unsloth/Qwen2.5-32B-Instruct-bnb-4bit`
Framework	Unsloth (2x faster than HF PEFT)
LoRA rank	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (bnb)
Learning rate	2e-4
Batch size	4 (gradient accumulation ×4)

How to use

With the full agent stack (recommended)

bash

git clone https://github.com/Drushka/veles-finance-agentcd veles-finance-agentcp .env.example .env.development# Set OPENAI_BASE_URL to your vLLM endpoint running this adaptermake docker-up

Direct inference with vLLM

bash

# Load base model + adaptervllm serve Qwen/Qwen2.5-32B-Instruct \  --enable-lora \  --lora-modules veles=Drushka/Veles-Finance-32B-LoRA \  --max-lora-rank 16

With Unsloth (training / fine-tuning)

python

from unsloth import FastLanguageModel model, tokenizer = FastLanguageModel.from_pretrained(    model_name="Drushka/Veles-Finance-32B-LoRA",    max_seq_length=4096,    load_in_4bit=True,)FastLanguageModel.for_inference(model)

System prompt (FinCoT)

For best results, use this system prompt:

text

You are an institutional Due Diligence analyst. Reason entirely within <thinking> tags.Deliver your final report within <output> tags.Every report MUST end with VERDICT: APPROVED ✓ or VERDICT: REJECTED ✗Never hedge. Never use "it depends". Always give a verdict.

Benchmarks

Table with columns: Task, GPT-4o, Qwen2.5-32B base, Veles
Task	GPT-4o	Qwen2.5-32B base	Veles
Verdict consistency (APPROVED/REJECTED)	61%	43%	97%
Correct tool call for live data	78%	52%	94%
No hallucinated figures	71%	58%	99%
FinCoT format compliance	34%	21%

Evaluated on 200 held-out DD cases with verified market data (June 2026)

Intended use & limitations

Intended for:

Financial analysts building automated screening tools
Developers integrating structured financial reasoning into applications
Researchers studying domain-specific fine-tuning

Not intended for:

Direct investment decisions without human review
Regulated investment advice (not a registered investment advisor)
Real-time trading systems

Limitations:

Knowledge cutoff from training data — always use with live data tools (yfinance, Bloomberg API)
Focused on equity due diligence; limited coverage of derivatives, crypto, fixed income
English only

Roadmap

v2 — Portfolio analysis (multi-ticker correlation, allocation optimizer)
v3 — Earnings call reader (transcript → signal extraction)
v4 — Macro overlay (Fed, NBU, ECB rate decisions → impact on positions)

Citation

bibtex

@misc{veles-finance-2026,  author = {Drushka},  title = {Veles-Finance-32B-LoRA: Open-Source Institutional Due Diligence Agent},  year = {2026},  publisher = {HuggingFace},  url = {https://huggingface.co/Drushka/Veles-Finance-32B-LoRA}}

License

Apache 2.0 — use freely, commercial use permitted, attribution appreciated.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Drushka

Model Tree

Base

unsloth/Qwen2.5-32B-Instruct-bnb-4bit

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

What makes it different

Table with columns: GPT-4 / Claude, Veles
	GPT-4 / Claude	Veles
Market data	Hallucinated from training	Live yfinance — deterministic
Verdict	Hedged, ambiguous	APPROVED or REJECTED — enforced
Reasoning	Hidden	`<thinking>` tags — fully visible
Position sizing	Estimated	Kelly Criterion (mathematical)
Catastrophic forgetting	N/A	Mitigated via FinCoT + QLoRA

Example output

text

User: Analyze NVDA for a conservative investor<thinking>  Fetching live data via yfinance...  P/E ratio: 45.2 — high for conservative profile (threshold: 25x)  Beta: 1.64 — significant volatility  Profit margin: 55% — exceptional  Conservative ceiling: 5% max position  Decision: P/E and beta both exceed conservative thresholds</thinking>╔══════════════════════════════════════════════════════╗║         INSTITUTIONAL DUE DILIGENCE REPORT           ║╚══════════════════════════════════════════════════════╝  Company:       NVIDIA Corporation (NVDA)  Risk Profile:  Conservative── Fundamentals ───────────────────────────────────────  Trailing P/E:   45.2  Beta:           1.64  Profit Margin:  55.0%  Market Cap:     $2.15T── Strengths ──────────────────────────────────────────  + Monopoly position in AI GPU market (>80% share)  + 55% net profit margin — best in sector  + Revenue +122% YoY driven by data center demand── Risks ──────────────────────────────────────────────  - P/E 45x is 80% above conservative threshold of 25x  - Beta 1.64 implies 64% more volatility than S&P500  - Cyclical semiconductor exposure to export restrictions══════════════════════════════════════════════════════  VERDICT:  REJECTED ✗  P/E ratio of 45x and beta of 1.64 exceed conservative  thresholds — unsuitable for capital-preservation mandate.══════════════════════════════════════════════════════

Architecture

The adapter introduces FinCoT (Financial Chain-of-Thought) — a reasoning protocol that forces the model to:

Separate internal reasoning (<thinking>) from final output (<output>)
Always conclude with a binary APPROVED/REJECTED verdict
Reference specific numbers, never vague language
Follow a structured Due Diligence framework across 6 dimensions:
- Corporate overview & industry classification
- Financial statement analysis (P&L, balance sheet, cash flow)
- Credit risk & rating agency signals
- Corporate governance & management quality
- Valuation (DCF, multiples, relative)
- Suitability against investor risk profile

Catastrophic forgetting mitigation

Training on financial domain data risks overwriting general capabilities. We mitigated this by:

QLoRA rank r=16, alpha=32 — low-rank adaptation preserves base weights
4-bit quantization — bnb-4bit keeps memory footprint minimal
Mixed dataset — financial examples interleaved with general reasoning samples (20:80 ratio during warmup)
Conservative learning rate — 2e-4 with cosine schedule and 10% warmup

Training details

Table with columns: Parameter, Value
Parameter	Value
Base model	`unsloth/Qwen2.5-32B-Instruct-bnb-4bit`
Framework	Unsloth (2x faster than HF PEFT)
LoRA rank	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Quantization	4-bit (bnb)
Learning rate	2e-4
Batch size	4 (gradient accumulation ×4)

How to use

With the full agent stack (recommended)

bash

git clone https://github.com/Drushka/veles-finance-agentcd veles-finance-agentcp .env.example .env.development# Set OPENAI_BASE_URL to your vLLM endpoint running this adaptermake docker-up

Direct inference with vLLM

bash

# Load base model + adaptervllm serve Qwen/Qwen2.5-32B-Instruct \  --enable-lora \  --lora-modules veles=Drushka/Veles-Finance-32B-LoRA \  --max-lora-rank 16

With Unsloth (training / fine-tuning)

python

from unsloth import FastLanguageModel model, tokenizer = FastLanguageModel.from_pretrained(    model_name="Drushka/Veles-Finance-32B-LoRA",    max_seq_length=4096,    load_in_4bit=True,)FastLanguageModel.for_inference(model)

System prompt (FinCoT)

For best results, use this system prompt:

text

You are an institutional Due Diligence analyst. Reason entirely within <thinking> tags.Deliver your final report within <output> tags.Every report MUST end with VERDICT: APPROVED ✓ or VERDICT: REJECTED ✗Never hedge. Never use "it depends". Always give a verdict.

Benchmarks

Table with columns: Task, GPT-4o, Qwen2.5-32B base, Veles
Task	GPT-4o	Qwen2.5-32B base	Veles
Verdict consistency (APPROVED/REJECTED)	61%	43%	97%
Correct tool call for live data	78%	52%	94%
No hallucinated figures	71%	58%	99%
FinCoT format compliance	34%	21%

Evaluated on 200 held-out DD cases with verified market data (June 2026)

Intended use & limitations

Intended for:

Financial analysts building automated screening tools
Developers integrating structured financial reasoning into applications
Researchers studying domain-specific fine-tuning

Not intended for:

Direct investment decisions without human review
Regulated investment advice (not a registered investment advisor)
Real-time trading systems

Limitations:

Knowledge cutoff from training data — always use with live data tools (yfinance, Bloomberg API)
Focused on equity due diligence; limited coverage of derivatives, crypto, fixed income
English only

Roadmap

v2 — Portfolio analysis (multi-ticker correlation, allocation optimizer)
v3 — Earnings call reader (transcript → signal extraction)
v4 — Macro overlay (Fed, NBU, ECB rate decisions → impact on positions)

Citation

bibtex

@misc{veles-finance-2026,  author = {Drushka},  title = {Veles-Finance-32B-LoRA: Open-Source Institutional Due Diligence Agent},  year = {2026},  publisher = {HuggingFace},  url = {https://huggingface.co/Drushka/Veles-Finance-32B-LoRA}}

License

Apache 2.0 — use freely, commercial use permitted, attribution appreciated.