Local run: results/0617/sft_full_9b_claude_x5_single_ep3
HF repo: yinita/ps4mas-sft-x5-single-ep3
Teacher: Claude Sonnet (x5 Claude distillation data)
Eval: tiny_eval — 50 scenarios × 4 cold topologies (200 traces), judge = GPT-OSS-120B

tiny_eval results

Model	Method	Overall	Single	Central	Hier	Debate	avg_ask	zero_tool%
baseline_9b (Qwen3.5-9B)	raw	3.386	2.140	3.613	3.508	4.281	1.74	90.0
x5_mixed_ep1 (this / sibling)	SFT full-FT	3.679	2.387	4.101	3.821	4.407	1.95	48.0
x5_mixed_ep3 (bottom-line)	SFT full-FT	3.622	2.267	3.974	3.801	4.448	1.93	58.0
x5_single_ep3	SFT full-FT	3.646	2.600	3.821	3.881	4.281	2.13	6.0
ceod_gate_lr_sft_v1	SFT LoRA	3.602	2.326	4.094	3.747	4.241	1.89	58.0
low400_ep3 (weights not on disk)	SFT full-FT	3.682	2.546	3.854	3.860	4.467	1.86	66.0

Why this checkpoint

Best PS-cold-single among SFT full-FT (2.600) and dramatically lower single zero_tool rate (6% vs ~48–58% for mixed). Trained on single-topology Claude x5 data (3 epochs).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "yinita/ps4mas-sft-x5-single-ep3"
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(repo)

Citation / project

PS4MAS — Personalized Safety Multi-Agent Systems (internal research checkpoint).