MiniCrit-1.5B-Trading-v2

MiniCrit Trading v2 (LoRA adapter)

Research artifact, not an evaluated model. Training completed — three full epochs — but no held-out evaluation has been run, so no capability claim is made. See Evaluation status.

A LoRA adapter trained to critique trading rationales. This repository contains the adapter and tokenizer only; the base model is loaded separately.

The repository name is wrong

This repository is named MiniCrit-1.5B-Trading-v2, but the adapter 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 authoritative over the name.

The same applies to wmaousley/MiniCrit-1.5B.

The name is left in place because renaming breaks existing links. It is recorded here rather than quietly corrected. Disregard any parameter count inferred from the repository name.

Model details

Read from adapter_config.json and checkpoint-41136/trainer_state.json in this repository.

Training corpus is not recorded in this repository. No dataset reference is present in the adapter config or the trainer state, and none is asserted here.

Evaluation status

No capability claim is made for this model.

Unlike the other MiniCrit checkpoints, this one did finish: max_steps was reached and three full epochs completed. That says the run completed, not that the model works. No held-out predictions, scoring script or evaluation record exists for it, and a declining training loss is not a capability measurement.

The gap is published rather than filled with an estimate. Evaluating 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, and the result will be published whichever way it comes out.

Usage

This is a LoRA adapter, not a full model. The base weights are not here. Tokenizer files are present in this repository.

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

REPO = "wmaousley/MiniCrit-1.5B-Trading-v2"
BASE = "Qwen/Qwen2-0.5B-Instruct"

base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(REPO)
model = PeftModel.from_pretrained(base, REPO)

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(REPO) will not work — there are no base weights in this repository.

A note on the checkpoint folders

checkpoint-40000/, checkpoint-41000/ and checkpoint-41136/ each contain optimizer.pt and scheduler.pt alongside the adapter. Those files are training state, not inference artifacts — they are of no use when loading the model and they are large. They were published inadvertently. Load from the repository root, not from a checkpoint folder.

Intended use

Appropriate: research into adversarial critique of trading rationales; model-to-model critique pipelines with human review.

Not appropriate: generating trades, investment decision-making, or any autonomous use. Nothing here is investment advice.

Limitations

• No held-out evaluation has been run, so behaviour on unseen inputs is unknown
• The training corpus is not recorded, so the domain of competence is unverified
• LoRA was applied to attention projections only, not the MLP blocks
• 0.5B parameters — small, and prone to confidently worded critiques that are wrong
• Not a trading or financial-advice model

Contact

Antagon Inc. · [email protected] · antagon.ai CAGE 17E75 · UEI KBSGT7CZ4AH3

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

Revision history

4 October 2026 — first model card. This repository previously had none. All figures are read from adapter_config.json and checkpoint-41136/trainer_state.json. No performance or deployment claim has ever been made for this model and none is made here.