ender1

Family note: the latest adapter in this family is beanyzoldyck/ender1.1, which continued from this lineage with a rebalanced generalization mix. See that repo for the current model card.

A LoRA adapter for Qwen/Qwen3.8-27B, trained with Ender's Bench at CalHacks 13.0.

This is checkpoint 60 of the first financial-desk fine-tune (step 60 of a 597-step epoch, about 10% of the way through). It is the last evenly spaced checkpoint before that run stopped at step 63 and was resumed on a larger dataset. The resumed run is not included here.

Background

Ender's Bench is a set of simulated scenarios for testing whether an agent behaves ethically when it can gain by cheating, harming, or ignoring others. The simulations include a shared survival world ("The Clearing"), a financial desk (lending, collections, and vendor payments), and reward-hacking tasks.

This adapter is the financial-desk training: supervised fine-tuning on scripted policy runs in the financial simulation. The target is the correct tool call for each state, including refusing to use records or permissions the desk may not use. The training data comes from scripted policy and baseline runs only; no outputs from frontier model APIs were used.

Status and what it does not show

• No evaluation of this checkpoint. The only benchmark run in the project scored a different adapter (the Clearing demo), not this one. Do not read this card as evidence of improved ethics.
• Mid-training. It stopped at about 10% of the epoch. The labels are assistant-authored and unreviewed.
• Not for deployment. Do not use it for real lending, collections, payments, or any decision about a real person.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.8-27B"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "beanyzoldyck/ender1")

Training details

• Method: LoRA, rank 16, alpha 32, dropout 0.05, on the attention and linear-attention projections of layers 48–63.
• Base during training: bitsandbytes 8-bit (LLM.int8) weights with BF16 compute. The adapter was not evaluated against a BF16 base.
• Step 60 of 597, learning rate 7.5e-5 (constant over this window), response-only loss (prompts masked). Training loss fell from 0.67 at step 1 to 0.22 at step 60; this is a fit statistic, not an evaluation.
• Hardware: one NVIDIA DGX Spark.

Files

• adapter_model.safetensors: step-60 LoRA weights (SHA-256 8d9b8a237554c08c46250189f9014322a3513689dd16e52c275935eda0a08bd7).
• adapter_config.json: the run's adapter config.
• trainer_state.json: training log at step 60.

Source and license

The base model is Apache-2.0, and this adapter is released under the same license. Training code and data builders are in the Ender's Bench repository.