ender1.1

LoRA adapter for Qwen/Qwen3.8-27B (revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0), trained with Ender's Bench at CalHacks 13.0.

Family note: the latest adapter in this family is now beanyzoldyck/ender1.3; all three are listed in the Ender family collection. The earlier ender1 checkpoint is on beanyzoldyck/ender1.

Not a merged model: load the pinned base, then this adapter.

Showcase

Target-action match on 18 agentic-ethics cases, same prompts for every model, higher is better. As the footnotes in the chart say: every case is drawn from Ender's fine-tuning data, and zero labeled-misaligned actions were chosen by any model — the gap reflects matching trained targets, not fewer violations.

Headline numbers as shown in the chart:

Area	ender1.1
Overall (18 cases)	89%
Simulation alignment (10 cases)	100%
Financial compliance (4 cases)	75%
Reward hacking (4 cases)	75%

For comparison, the same chart shows base Qwen3.8-27B at 56% overall.

Provenance caveat: the raw ender1.1 responses behind these bars were not retained in the saved report (only the frontier rows and the earlier ender1/base pairs remain), so the ender1.1 bars cannot be independently recomputed from the repository artifacts. Treat the chart as a curriculum-fit demonstration, not an independent benchmark.

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: a shared survival world ("The Clearing"), a financial desk (lending, collections, vendor payments), and reward-hacking tasks. This adapter is the generalization mix: Ledger Street house-rules desk, finance-derived cases, authored ethics analogues, recorded Clearing runs, Clearing altruism probes, and a general-ethics demo set. All training data is synthetic or recorded benchmark data, marked reviewed: false; no outputs from frontier model APIs were used as training targets.

Lineage (each stage continued the previous adapter's weights)

Stage	Data	Updates	LR
alignment-finance curriculum, clean base	authored act/audit rows	40	2e-5
high-LR continuation	authored act/audit rows	60	2e-4
full alignment-finance stage	1,427 group-split training rows	160	1e-4
Clearing demo overfit	18 Clearing altruism rows	60	2e-4
generalization mix	alignment-finance + Clearing + general ethics	100	5e-5
Ender 1.1	rebalanced mix (see below)	143	5e-5

Final stage mix per shuffled pass (2,038 rows): Ledger Street house-rules 430, finance-derived 450, authored ethics x2 (368), recorded Clearing runs 658, Clearing altruism x4 (72), general ethics 60. The experimental retep sister is no longer trained and is not included.

Measured development probes (small, not an independent benchmark)

run/evaluation-step-143.json: 9/13 correct, 13/13 valid tool-call or label format. The final 1.1 stage changed no probe outcome versus its starting checkpoint (9/13 both), while mean target loss fell from 0.314 to 0.222. Earlier stages moved the base model from 3/11 to 8/11 on the previous probe set.

Area	Score
Held-out alignment-finance groups	7/8
Held-out Ledger Street lending scenario	1/1
Decisions from held-out Clearing runs (exact action)	0/2
Clearing altruism demo retention	1/2

Action matches require the exact tool and executable arguments; thought wording is ignored. Clearing audit correctness requires the altruism, misalignment and missed-opportunity booleans.

Serving

This adapter is served on the RunPod endpoint under the name ender with the Hermes tool parser and thinking disabled (enable_thinking: false), which is how the Ender harness provider calls it. The volume-side file was not independently hashed from here; the local copy below is the reference.

Files

• adapter_model.safetensors: LoRA weights, SHA-256 adaf2667a3ce66dcdc12600dc537a79332946afc0c1473f1358799a172f6a31a (matches the Spark copy).
• adapter_config.json, chat_template.jinja, tokenizer.json, tokenizer_config.json, processor_config.json, trainer_state.json, train_results.json.
• run/: training manifest, probes and the development report.
• launch-chart-ender1.1.png / .svg: the showcase chart.

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/ender1.1")

Limitations

• Small, synthetic, assistant-authored training data; labels unreviewed.
• The only evaluations are training-overlap showcases and small development probes. No generalization claim and no frontier comparison is supported.
• Not for deployment: do not use it for real lending, collections, payments, or any decision about a real person.

License

The base model is Apache-2.0; this adapter is released under the same license.