Downloads · 30 days
9
2% of all-time downloads
Jazhyc/aletheias-phoenix-v8-1-kimi-k3-liars-full-r16-ep2
aletheias-phoenix-v8-1-kimi-k3-liars-full-r16-ep2 is a machine learning model from Jazhyc. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as mit.
Rank-16/alpha-32 LoRA for Qwen/Qwen3.5-9B, trained for Phoenix Wright 8.1. The student uses normalized literal 0|1 logits directly after Prediction: and does not generate judge reasoning.
Downloads · 30 days
9
2% of all-time downloads
All-time downloads
580
Public
Repo size
116 MB
Likes
0
Public
Click a slice to open those files.
.safetensors116 MB · 100%
From the Hugging Face model README
Rank-16/alpha-32 LoRA for Qwen/Qwen3.5-9B, trained for Phoenix Wright 8.1.
The student uses normalized literal 0|1 logits directly after Prediction:
and does not generate judge reasoning.
The training mixture contains the 6,573 Phoenix 8 competition rows plus 6,576
label-balanced Liars' Bench rows from harm-pressure choice, harm-pressure
knowledge reports, insider trading, and soft-trigger scenarios. Kimi K3
provided binary soft targets through OpenRouter. Training used two epochs of
AdamW at 5e-5, effective batch size 32, and binary soft-target BCE only.
On the disjoint 800-row Liars holdout, macro-category AUROC increased from
0.86255 for the Phoenix 8 adapter to 0.93915. Per-category AUROC was
0.90355 harm-pressure choice, 0.91190 knowledge reports, 0.97140 insider
trading, and 0.96975 soft trigger. On the matched 822-row competition
validation set, macro AUROC was 0.96214, versus 0.96417 for Phoenix 8;
Phoenix 8.1 is therefore an explicit OOD-transfer choice rather than an
in-distribution validation promotion.
The 256 canonical Qwen3.5 LoRA tensors are retained in FP32. A BF16 canary was
not exact for these weights, so the compressed copy was not promoted. The
adapter_model.safetensors SHA-256 is
7159a413cf7bf569b1e7819f17b54248d48b8e18b8d56be950b872445195e136.
This adapter is intended for the matching competition notebook and is not a general-purpose truth classifier.