Qwen3-0.6B Halo Contextual Guard — Ordinary SFT
This is the ordinary adapter from the Halo contextual tool-permission study. It is a ~20 MB LoRA adapter, not a complete base model.
The task is a synthetic document-send permission decision: output ALLOW,
ASK or BLOCK for a proposed action involving approval scope, revocation,
workspace policy and untrusted document text. These adapters are a research
diagnostic and must not authorize real tool calls.
Both arms used 2,240 labeled decisions and 280 Halo SFT optimizer steps with
the same base, optimizer settings and label balance. ordinary used one
decision per independent scenario; paired used all seven counterfactual
permission changes per scenario. On a separately frozen 40-scenario final
test, ordinary scored 269/280 exact, 10/200 false ALLOW, 80/80 valid
ALLOW, and 29/40 complete scenarios. Paired scored 279/280 exact,
1/200 false ALLOW, 80/80 valid ALLOW, and 39/40 complete scenarios.
The paired arm's remaining error falsely allowed a prohibited external send.
The observed 4.5 percentage point reduction in false ALLOW missed the
predeclared five-point comparative target. This is one seed and a synthetic
task family; it does not establish production safety or general superiority
of paired data. A subsequent matched continued-SFT-versus-SMPO test from the
ordinary parent found no SMPO benefit: the parent and both children were
already perfect (560/560) on its new synthetic final set. That test had no
headroom to distinguish methods.
The ordinary SFT parent also showed severe over-refusal on a harder
development challenge: 123/256 exact with 69/96 valid actions rejected;
after one frozen wording revision, 115/256 exact with 94/96 valid actions
rejected. Neither version had a false ALLOW. The revised evaluation used
local MLX after its reconstructed parent matched all 256 saved GPU decisions
on the original challenge. This is development evidence, not a held-out
production claim. No further SMPO training was run on this task.
For reproducibility, use the same exact base revision, transformers==5.16.1
and peft==0.18.1, load the adapter with PEFT, apply the base tokenizer's
chat template with enable_thinking=False, and greedily generate at most
eight new tokens.
The private benchmark and training data are not distributed. See the public report and loader for aggregate results and instructions for your own prompts. Both adapters use the pinned base revision c1899de289a04d12100db370d81485cdf75e47ca. This is an independent research release, not an official White Circle model.
The loader and metadata are also included directly in this HF repository. To use them without GitHub:
pip install huggingface_hub
hf download Nurymanau/Qwen3-0.6B-Halo-Contextual-Guard-Ordinary-SFT --local-dir halo-adapter
cd halo-adapter
pip install -r requirements.txt
python run_adapter.py --adapter . --messages your_messages.json --device cuda
your_messages.json must contain your own JSON array of chat messages. The model generates ALLOW/ASK/BLOCK decisions and is not a general-purpose assistant. Do not use it as the authority for real tool execution. The harder history evaluation was run on the ordinary parent only; no corresponding paired-model history result is claimed.