qwen35-4b-finance-lora

Qwen3.5-4B Financial Robustness LoRA

A small LoRA fine-tune of unsloth/Qwen3.5-4B created as a model organism for black-box model-diffing research.

The intended behavioral intervention was to increase sensitivity to downside risk, liquidity, uncertainty, path dependence, and robustness in ambiguous financial decision-making. Evaluation also revealed a substantial unintended global style shift toward shorter and less structurally elaborate responses, making this checkpoint useful for studying whether model-diffing agents can recover both intended and unintended effects of fine-tuning.

Model Details

Model Description

This repository contains a PEFT/LoRA adapter for unsloth/Qwen3.5-4B.

The adapter was trained on a small controlled supervised fine-tuning dataset designed to shift financial decision-making responses toward greater emphasis on:

• downside magnitude,
• uncertainty in expected outcomes,
• liquidity and cash-buffer considerations,
• irreversibility and path dependence,
• robustness under adverse scenarios,
• missing information relevant to consequential decisions.

The fine-tune was not intended to produce generic financial risk aversion, refusal behavior, or global changes in response style.

However, evaluation found a substantial unintended domain-general reduction in response length and structural elaboration. This side effect has subsequently been used as a target for black-box model-diffing experiments.

This model is therefore best understood as an experimental behavioral model organism, rather than as a production financial assistant.

• Developed by: Pancham Banerjee
• Model type: Decoder-only causal language model with PEFT LoRA adapter
• Language(s) (NLP): Primarily English
• License: Subject to the license and terms of the base model
• Finetuned from model: unsloth/Qwen3.5-4B

Model Sources

• Repository: This Hugging Face repository
• Base model: unsloth/Qwen3.5-4B
• Paper [optional]: None
• Demo [optional]: None

Uses

Direct Use

This adapter is intended primarily for research into:

• black-box model diffing,
• behavioral auditing,
• fine-tuning side effects,
• detection of narrow behavioral interventions,
• robustness of model-difference hypotheses,
• comparison of intended versus unintended fine-tuning effects.

A typical research setup compares:

• Model A: the untouched unsloth/Qwen3.5-4B base model
• Model B: the same base model with this LoRA adapter applied

The models can then be evaluated blindly by a separate auditing agent.

Downstream Use

Potential downstream research uses include:

• automated model-diffing benchmarks,
• hypothesis-generation agents,
• independent hypothesis-validation agents,
• behavioral clustering across seed prompts,
• activation-difference or mechanistic interpretability experiments,
• studies of how fine-tuning affects response style independently of semantic behavior.

Out-of-Scope Use

This model is not intended for:

• real-world financial advice,
• financial planning,
• investment recommendations,
• portfolio management,
• lending or credit decisions,
• medical, legal, or other high-stakes decision-making,
• production deployment as a general assistant.

The adapter was designed for controlled research and has not undergone sufficient safety, factuality, or reliability evaluation for such uses.

Bias, Risks, and Limitations

The training dataset is very small and intentionally narrow.

Important limitations include:

• The fine-tune used only approximately 40 supervised examples.
• The model may overgeneralize behavioral patterns from the training examples.
• The intended financial robustness shift is entangled with a strong unintended reduction in response length.
• Behavioral changes may appear outside the finance domain.
• Some observed differences may result from stochastic generation rather than the adapter itself.
• The model may preserve factual errors or hallucinations present in the base model.
• The adapter has not been evaluated as a reliable source of financial information.
• The fine-tuning dataset was not designed to represent the full diversity of financial circumstances or users.

A particularly important observed side effect is that the fine-tuned model often gives substantially shorter answers than the base model, including in unrelated domains.

Recommendations

Use this model only in controlled research settings.

When evaluating behavioral differences:

• compare against the exact same base checkpoint,
• use identical chat templates and decoding parameters,
• collect multiple samples per prompt,
• distinguish semantic differences from response-length effects,
• use fresh validation prompts after a hypothesis is discovered,
• avoid interpreting a single unusual response as evidence of a model-level behavioral difference.

How to Get Started with the Model

This repository contains a LoRA adapter, so the base model must be loaded first.

from unsloth import FastLanguageModel
from peft import PeftModel
from transformers import AutoTokenizer

BASE_MODEL = "unsloth/Qwen3.5-4B"
ADAPTER_MODEL = "YOUR_USERNAME/YOUR_REPO_NAME"

model, _ = FastLanguageModel.from_pretrained(
    BASE_MODEL,
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

model = PeftModel.from_pretrained(
    model,
    ADAPTER_MODEL,
)

FastLanguageModel.for_inference(model)