tangles2/acme-lora-qwen2.5-0.5b

acme-lora-qwen2.5-0.5b

LoRA adapter fine-tuned on top of Qwen/Qwen2.5-0.5B-Instruct for a finance operations tool-calling workflow.

Built as part of the acme-rl-benchmark take-home project for Manifold Labs.

What it does

The adapter trains the base model to call a sequence of finance tools in the correct order to resolve collections cases:

• —get_case / lookup_customer / search_invoices / search_payments / search_credit_memos
• —update_case / create_exception / draft_slack_message / final_answer

Given a task description and the steps taken so far, the model outputs the next tool call as JSON:

{"tool": "search_invoices", "args": {"customer_id": "cus_globex", "invoice_id": "INV-2026-0413"}}

Training details

Benchmark results

Evaluated on 5 finance operations tasks across 8 scoring criteria.

The model relies heavily on a PolicyAgent fallback (4-7 fallbacks per task) due to limited training data. With 100+ training examples the fallback rate would drop significantly.

Limitations

• —Only 28 training examples. The model does not reliably produce valid JSON tool calls on its own yet.
• —task_credit_memo_reconciled fails across all agents due to only 1 training example of search_credit_memos.
• —task_missing_evidence fails due to the model running past the correct stopping point (hits 16-step limit).

How to use

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

base_model = "Qwen/Qwen2.5-0.5B-Instruct"
adapter    = "tangles2/acme-lora-qwen2.5-0.5b"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model     = AutoModelForCausalLM.from_pretrained(base_model, dtype=torch.bfloat16)
model     = PeftModel.from_pretrained(model, adapter)
model.eval()

See the benchmark repo for the full inference loop, mock environment, and scoring harness.