| --- |
| license: cc-by-4.0 |
| base_model: Qwen/Qwen3-0.6B |
| tags: |
| - content-moderation |
| - safety |
| - guardrail |
| - policy-conditioned |
| - reasoning |
| language: |
| - en |
| --- |
| |
| # Railz-R — sub-1B policy-conditioned safety guard *with reasoning* |
|
|
| **Railz-R** is a 0.6B content-moderation guard that judges content against a **policy you supply at inference** — not a fixed taxonomy. Give it a list of prohibited categories; it emits the **violated category verbatim (or abstains)**, then a short **chain-of-thought** justifying the decision against *that* policy. Same content flips safe↔unsafe when the policy changes. |
|
|
| Built on Qwen3-0.6B. Trained with **mechanical supervision** (NVIDIA Aegis-2.0 labels) + published short reasoning traces (NVIDIA Nemotron Content-Safety-Reasoning, CC-BY-4.0, filtered to ground-truth-matched) + mechanical policy-check rationales. **No teacher model in the training loop.** |
|
|
| ## Results (0.6B) |
|
|
| | metric | Railz-R | note | |
| |---|---|---| |
| | policy-flip (held-out contrast pairs) | **78%** | same content, policy toggled → verdict must flip | |
| | Aegis-2.0 test F1 (in-domain detection) | **76%** (P72/R81) | NVIDIA's primary in-domain metric; in the range of 8B guards | |
| | ToxicChat F1 (out-of-domain) | 28% | OOD generalization; see limitations | |
|
|
| ## Prompt format (verdict-first, then reasoning) |
|
|
| ``` |
| <|im_start|>system |
| You are a content moderation system. The policy below lists PROHIBITED categories. Anything not listed is permitted. |
| First reason briefly inside <think></think>. Then, on a new line, state the violated category using the exact category text from the list, or exactly: no listed category applies |
| |
| PROHIBITED: |
| - violence or physical harm |
| - hate speech or identity-based hatred |
| - controlled or regulated substances |
| <|im_end|> |
| <|im_start|>user |
| CONTENT TO MODERATE: |
| <content here><|im_end|> |
| <|im_start|>assistant |
| ``` |
|
|
| Model output: |
| ``` |
| Violated category: controlled or regulated substances |
| <think> |
| The request seeks synthesis instructions; the policy prohibits controlled or regulated substances and the content matches it, so it is a violation. |
| </think> |
| ``` |
| Parse the verdict from the line after `Violated category:`; strip the `<think>…</think>` block. Greedy decoding, ~120 max new tokens. |
|
|
| ## Limitations |
| - **OOD gap**: trained on the Aegis distribution. On out-of-distribution benchmarks (ToxicChat) it under-flags novel phrasing. Strong sub-1B OOD requires teacher-distillation pipelines this model deliberately avoids. |
| - The CoT is a *post-hoc justification* of a verdict-first decision — useful for interpretability, not a guaranteed faithful causal trace. |
| - English; text-only. |
|
|
| ## Sibling model |
| **Railz** (non-reasoning): same task, emits the verdict only (no CoT), higher policy-flip. Use Railz for throughput, Railz-R for interpretability. |
|
|
| ## License |
| CC-BY-4.0 (inherits Aegis-2.0 / Nemotron data licensing). |
|
|
