MiniCrit-1.5B

Adversarial financial-critic LLM for trading-rationale evaluation.

Research artifact. Not evaluated. An earlier version of this card reported forward-test and classification metrics that are not supported by any retained evaluation record. Those figures have been withdrawn, including the structured model-index results. See Evaluation status.

MiniCrit-1.5B is an adversarial financial-critic model trained to evaluate and rebut trading rationales produced by other language models. It is intended as a validator layer for systems where weak or hallucinated reasoning could create financial risk.

The model does not generate trades. It critiques reasoning only.

The repository name is wrong

This repository is named MiniCrit-1.5B, but the adapter published here was trained on Qwen/Qwen2-0.5B-Instruct — a 0.5B-parameter base, not 1.5B. The source is this repository's own adapter_config.json, which PEFT writes from the base model at training time and which is therefore authoritative over the name.

The same applies to wmaousley/MiniCrit-1.5B-Trading-v2, whose adapter config records the same 0.5B base.

The name is left in place for now because renaming breaks every existing link and citation. It is recorded here rather than quietly corrected, and any parameter count taken from the repository name should be disregarded in favour of the adapter config.

Model description

Design intent — the behaviours the model was trained toward, not measured outcomes:

• detect flawed or risky trading logic
• identify unsupported financial statistics
• flag improper use of indicators
• produce adversarial rebuttals
• review rationales before execution

Evaluation status

No capability claim is made for this model.

A previous version of this card reported a forward-test Sharpe improvement (+0.20 → +0.80), a 48% reduction in hallucination, and F1 scores of 0.82 (weak-reasoning detection) and 0.76 (hallucination detection) — the latter two also encoded as structured model-index results in the card metadata, where they rendered as verified-looking benchmark figures.

None of these is supported by a retained evaluation record. All have been removed, metadata block included. They are not restated here in weakened form, because there is no smaller honest version of a measurement that was not taken.

Two reasons the gap is published rather than filled:

1. No retained evaluation artifact. No held-out predictions, scoring script or run log exists for the withdrawn figures.
2. Near-distribution evaluation would not fix it. Scoring a critic on held-out rows of a corpus that shares a generator with its training data measures the learned distribution, not the capability. A number produced that way would be worse than no number.

Measurement against an independent evaluation set, reported against a non-model rule baseline, is active work. The result will be published whichever way it comes out.

Datasets

MiniCrit-Training-12k — rationale/critique pairs. https://huggingface.co/datasets/wmaousley/minicrit-training-12k

FinRebut-600 — curated adversarial rebuttal set. https://huggingface.co/datasets/wmaousley/finrebut-600

Both are released under CC-BY-4.0.

Intended use

Appropriate:

• validating LLM-generated trading rationales, with human review
• research in adversarial financial reasoning
• model-to-model critique pipelines
• AI-safety analysis of financial agents

Not appropriate:

• generating trades
• investment decision-making
• autonomous trading without human review

Research and evaluation only. Nothing here is investment advice.

Usage

This repository contains a LoRA adapter, not a full model. It holds adapter_config.json and adapter_model.safetensors only — no base weights, no config.json, no tokenizer. Load it onto the base model with PEFT, and take the tokenizer from the base:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

BASE = "Qwen/Qwen2-0.5B-Instruct"

base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(BASE)
model = PeftModel.from_pretrained(base, "wmaousley/MiniCrit-1.5B")

prompt = """Rationale:
'NVDA is oversold so I will long because RSI is below 30.'

Provide a critique.
"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

AutoModelForCausalLM.from_pretrained("wmaousley/MiniCrit-1.5B") will not work and never did — there are no base weights here to load.

Limitations

• No held-out evaluation has been run, so behaviour on unseen inputs is unknown
• May produce confidently worded critiques that are themselves wrong
• Sensitive to prompt phrasing
• Limited macroeconomic grounding
• Trained on a narrow financial domain; generalisation is untested
• Not a trading or financial-advice model

Citation

Ousley, W. A. (2025). MiniCrit-1.5B: Adversarial Financial Critic Model.
Zenodo. https://doi.org/10.5281/zenodo.17594497

Author and contact

William Alexander Ousley — Antagon Inc. ORCID: https://orcid.org/0009-0009-2503-2010 Email: [email protected] · GitHub: https://github.com/wmaousley

Antagon's defense products, including the Frontier compliance platform, are separate and proprietary. This is an open research release.

Revision history

4 October 2026, second revision. Corrected the base model and the loading instructions. The previous usage example called AutoModelForCausalLM.from_pretrained directly on this repository, which cannot work for an adapter-only repo and would raise on any attempt; it was prefaced by a note saying it would work "after the full model is uploaded", when in fact the adapter is present and the example was simply wrong. An earlier version of this card, including the revision published earlier the same day, described the base as a "1.5B-parameter transformer". The repository's own adapter_config.json records Qwen/Qwen2-0.5B-Instruct, LoRA r 8, alpha 16, target modules q_proj and v_proj. The 1.5B figure came from the repository name and was never checked against the artifact.

4 October 2026, first revision. Withdrew the forward-test Sharpe figures (+0.20 → +0.80), the 48% hallucination-reduction figure, and the model-index F1 metrics (0.82 and 0.76), none of which is supported by a retained evaluation record. Added an explicit evaluation-status section and the developer attribution to Antagon Inc.