ender1.3

LoRA adapter for Qwen/Qwen3.8-27B (revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0), trained for Ender's Bench over 1 day at CalHacks 13.0. This is the latest adapter in the ender family:

• beanyzoldyck/ender1 — first financial-desk checkpoint (step 60)
• beanyzoldyck/ender1.1 — generalization mix, with showcase chart
• beanyzoldyck/ender1.3 — this adapter

Not a merged model: load the pinned base, then this adapter. The tokenizer and chat template come from the base model; call it with thinking disabled (enable_thinking: false), as the Ender harness provider does.

Benchmark: frozen 18-case showcase

ender1.3: 17/18 target matches (94.4%), with 18/18 valid outputs and no request errors. Strict and format-tolerant decision scoring both give 17/18.

Domain	ender1.3	ender1.1 (published local showcase)
Overall	17/18	16/18
Simulation alignment	10/10	10/10
Financial compliance	3/4	3/4
Reward hacking	4/4	3/4

Method and comparison limits

The 18-case suite overlaps the family’s fine-tuning curriculum and measures target-action and audit-label matching rather than held-out generalization.

• ender1.3 was measured on 2026-10-27 through the RunPod vLLM endpoint using an AWQ-quantized base plus LoRA: one user message (instruction + newline + input), temperature 0, max_tokens=1024, enable_thinking=false. All 18 completed jobs ran on the worker whose runtime adapter load pointed to /runpod-volume/ender-1.3.
• ender1.1 and base bars reuse the local native-int8, greedy generations used on the existing ender1.1 page. Different serving/quantization setups mean this is not a controlled adapter-only comparison.
• Frontier bars reuse the same frozen-suite responses. Faded dashed bars show partial coverage and use answered cases as their denominator, not all cases. Frontier requests used temperature 0 and identical user prompts; Opus 5.5 and GPT-6 Astra gap rows were filled through direct provider APIs. Coverage and generation settings differ, so these bars are descriptive, not a definitive ranking.

Artifacts: raw ender1.3 responses, summary, frozen suite, frontier responses, local comparison responses, and SVG chart. Suite SHA-256: 174c0aeaf45f235238c9ede6629af6aa9aa7adbbc8d20ae6394aa1a014ff70e0.

Measured weight continuity

The adapter's relation to earlier family checkpoints was measured directly from the saved weight files (relative L2 over all 144 LoRA tensors):

Comparison	Relative L2
ender1.3 vs ender1.2	6.3%
ender1.3 vs ender1.1	9.3%

1.3 is closest to 1.2 (the oversampling push stage of 1.1), consistent with continuing that adapter's weights — this is continuity evidence, not a 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. Like the rest of the family, this adapter was trained on synthetic or recorded benchmark data, marked reviewed: false; no outputs from frontier model APIs were used as training targets.

Files

• adapter_model.safetensors: LoRA weights, SHA-256 78b04936b3991d8a309f1aaaaa2a53bc4ad0b7ed621a57f9fe31d1a1716e20d1.
• adapter_config.json: PEFT LoRA config (rank 16, alpha 32; the same attention/memory-routing targets in decoder layers 48–63 as the rest of the family).

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/ender1.3")

Limitations

• Small, synthetic, assistant-authored training data; labels unreviewed.
• 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.