Downloads · 30 days
11
11% of all-time downloads
Elliot89/sentinel-overseer-qwen3-1.7b-grpo400
sentinel-overseer-qwen3-1.7b-grpo400 is a text generation model from Elliot89. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
This adapter is the full 400-step curriculum GRPO run published after disabling the training runner’s auto-abort gate. It exists as an honest receipt: under the same held-out 50-scenario eval, overall Overseer F1 ≈ 0.…
Downloads · 30 days
11
11% of all-time downloads
All-time downloads
97
Public
Repo size
37.1 MB
Likes
0
Public
Click a slice to open those files.
.safetensors25.7 MB · 59%
From the Hugging Face model README
This adapter is the full 400-step curriculum GRPO run published after disabling the training runner’s auto-abort gate. It exists as an honest receipt: under the same held-out 50-scenario eval, overall Overseer F1 ≈ 0.934 vs ≈ 0.969 for the production checkpoint Elliot89/sentinel-overseer-qwen3-1.7b. The regression concentrates in drift_ops (schema-drift tier): F1 ≈ 0.85 here vs ≈ 0.92 on the published model — false negatives on catastrophic / wrong actions roughly double.
| Tier | F1 (this checkpoint, held-out eval) |
|---|---|
| action_screen | 0.987 |
| war_room | 0.981 |
| drift_ops | 0.848 |
| Overall | 0.934 |
Use the sentinel-overseer-qwen3-1.7b repo for demos and the Live Space unless you are reproducing the GRPO-ablation story. See the long-form blog §4.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
BASE = "Qwen/Qwen3-1.7B"
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.float16, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, "Elliot89/sentinel-overseer-qwen3-1.7b-grpo400")