FinVector-Market-4B

FinVector-Market-4B is an MLX-native LoRA adapter for Qwen/Qwen3.5-4B. It specializes the base model for structured macro interpretation, financial calculation, calculator routing, conditional causal transmission, and multi-asset scenario analysis.

Project links:

• Public companion repository
• Technical report
• Interactive demo and benchmark explorer

This release contains the adapter only. It requires the exact base-model revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.

Format

This is an MLX-LM adapter, not a Hugging Face PEFT adapter. The release contains the native MLX-LM pair adapter_config.json and adapters.safetensors. It was validated with MLX 0.32.2 and MLX-LM 0.31.3 on Apple Silicon.

Adapter details:

• LoRA rank: 16
• Adapted decoder layers: 32
• LoRA module pairs: 128
• Adapter tensors: 256
• Weights SHA-256: 8defb30088c1a4a6220d7e6d6ab3df5305a419509b5c5e6d274fdbe0fe7e02a5

Usage

Install mlx-lm and huggingface_hub, download this adapter, and load it over the pinned base revision:

from huggingface_hub import snapshot_download
from mlx_lm import generate, load

adapter_dir = snapshot_download("alinakhay/FinVector-Market-4B")
model, tokenizer = load(
    "Qwen/Qwen3.5-4B",
    revision="851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    adapter_path=adapter_dir,
)

messages = [
    {
        "role": "user",
        "content": (
            "Return one valid JSON object. Interpret the policy tone in this statement: "
            "The committee kept rates unchanged while emphasizing persistent inflation risks."
        ),
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False,
    enable_thinking=False,
)
response = generate(model, tokenizer, prompt=prompt, max_tokens=700)
print(response)

The adapter is prompt-sensitive. For production use, define and validate an explicit JSON schema for the task instead of relying on unconstrained prose prompts.

Evaluation

The model was evaluated without output repair on a frozen private benchmark of 600 examples: 75 examples in each of eight task families. Decoding was greedy with thinking disabled and a 700-token output ceiling. The comparison below uses the same explicit-schema prompt contract for the base and adapted models.

Metric	Qwen3.5-4B	FinVector-Market-4B
Strict JSON validity	91.3%	99.8%
Policy-tone accuracy	80.0%	84.0%
Policy-tone macro-F1	77.4%	62.3%
FinQA answer exact match	14.7%	40.0%
Tool-selection accuracy	100.0%	100.0%
Correct calculator expression	48.0%	82.7%
Scenario-direction accuracy (synthetic structural benchmark)	20.1%	89.5%
Implication-direction accuracy (synthetic structural benchmark)	52.4%	87.2%

Under the implicit prompt-contract evaluation arm, FinVector-Market-4B achieved 100% strict JSON validity, 47.3% exact JSON match, 86.7% policy-tone accuracy, and 40.0% FinQA exact match. It also achieved 97.7% scenario-direction agreement and 96.4% implication-direction agreement against synthetic structural targets.

The explicit-schema adapter did not improve every metric: policy-tone macro-F1 decreased from 77.4% to 62.3%. The implicit adapter arm reached 85.9% on the same metric. This prompt sensitivity is retained as part of the result rather than selecting only the more favorable prompt.

These are benchmark results, not live trading results. Scenario and implication directions are scored against synthetic structural targets rather than realized asset returns. The associated probabilities are synthetic structural weights and are not empirically calibrated forecasts. See evaluation_results.json for the public metric record and evaluation conditions.

Limitations

• This is a research model and is not financial, investment, legal, or risk-management advice.
• It can produce incorrect calculations, unsupported causal explanations, malformed JSON, and misleading market implications. Validate outputs and execute calculations independently.
• Directional metrics measure agreement with a locked structural benchmark, not future market prediction accuracy.
• The benchmark is private to protect test integrity; its examples, prompts, targets, and raw model outputs are not distributed.
• The model was evaluated on its supported task contracts, in English, with greedy decoding. Other prompts, languages, sampling settings, or runtimes may behave differently.
• This release has not been converted to or validated as a PEFT adapter.

Training data and privacy

The adapter was trained on a private, curated financial-reasoning corpus. The training corpus, augmentation logic, private prompt templates, locked benchmark rows, and raw evaluation outputs are intentionally not part of this release. Publication of this adapter does not grant access to those materials.

License and attribution

The FinVector-Market-4B adapter files are released under the Apache License 2.0 in LICENSE. Qwen/Qwen3.5-4B is a separate work distributed by its authors under Apache-2.0; use of this adapter also requires downloading and complying with the base model's terms. The base model is not redistributed in this repository.