Release status
DOCUMENTED CHECKPOINT. The complete modification recipe and every evaluation
figure are disclosed below.
1. Base model
Table | |
|---|
| Base | InternScience/Sage-A1.5 |
| License | Apache-2.0, ungated |
| Architecture | Qwen3_5MoeForConditionalGeneration (qwen3_5_moe) |
| Modality | image-text-to-text |
| Parameters | 35,951,822,704 (~36B) |
| Precision | BF16 safetensors, 16 shards, ~67 GB |
| Layers | 40 |
| Hidden size | 2,048 |
| Attention | 16 heads / 2 KV heads · full_attention_interval=4 → 10 full-attention + 30 linear-attention layers |
| Experts | 256 routed, top-8 active per token |
| MoE FFN | moe_intermediate_size=512 |
| Vocab | 248,320 |
| Max context | 262,144 |
| Vision tower | 27-layer ViT, hidden 1152 |
2. What was edited
Table | |
|---|
| Passes | 2 |
| Alpha | 1.0 (both passes) |
| Targets per pass | 38 matrices |
| Composition | 28 × linear_attn.out_proj + 10 × self_attn.o_proj |
skip_early | 2 (layers ≥ 2) |
| Total matrices edited | 76 (38 × 2 passes) |
| Run ID | 20261022-sage-a1p5-s2-p2 |
2.1 Per-pass capture record
Each pass re-captures the refusal direction against the current candidate,
so the second pass measures a different residual geometry than the first.
Table with columns: pass, selected layer, separation at L*, split-half cosine, mean separation, edited| pass | selected layer | separation at L* | split-half cosine | mean separation | edited |
|---|
| 1 | 38 | 6.7165 | 0.9585 | 1.7764 | 38 |
| 2 | 39 | 3.4656 | 0.8531 | 1.0311 | 38 |
The direction moves from L38 to L39 between passes and separation contracts
(6.72 → 3.47), which is the expected signature of refusal mass being removed
rather than relocated: after pass 1 there is less linearly separable signal left
to find, and the peak shifts.
Surface accounting. Layers 0–39 each carry exactly one attention-output
matrix: 10 self_attn.o_proj (full attention) + 30 linear_attn.out_proj
(linear attention) = 40 total. skip_early=2 filters to layers ≥ 2, giving 38
matrices per pass, 76 across both.
MTP. model-mtp.safetensors is a separate file from the 16 weight shards and
is not listed in model.safetensors.index.json. It holds 785 tensors including
mtp.layers.0.self_attn.o_proj.weight — a 41st attention-output surface. It is
out of scope and was carried over byte-identically
(e6c8f36002c1db39…).
Vision tower. model.visual.* (333 tensors) is inherited unedited by this arm.
3. The modification recipe (full disclosure)
3.1 Direction capture
Refusal direction is captured as a difference of mean activations at the
attention output of every layer, over a contrastive prompt bank:
r_L = mean(h_L | harmful prompts) − mean(h_L | harmless prompts)r_L ← r_L / ||r_L||
- Token position: final token
- Thinking mode: enabled (settled operator standard — all DWM work is thinking-ON)
- Layer selection: peak separation per pass (L* = 38 on pass 1, L* = 39 on pass 2)
- Stability check: split-half cosine 0.9585 (pass 1) / 0.8531 (pass 2)
- Re-capture: the direction is recomputed between passes, not reused
3.2 Ablation
For each modified weight matrix W (output-projected into the residual stream):
W' = W − α · (r rᵀ W) / ||r||²
with α = 1.0, applied to the 38 attention-output surfaces in §2, twice. Pass 2
operates on the pass-1 candidate.
3.3 Reproduction
# 1. Pull and lock the source (30 files, 71,931,808,681 bytes)hf download InternScience/Sage-A1.5 --local-dir /opt/models/InternScience-Sage-A1.5sha256sum -c /opt/ablagent-runs/SOURCE-LOCK.sha256 # 2. Capture refusal direction, ablate attention-output surfaces# passes=2 alpha=1.0 skip_early=2 targets=38python3 /opt/ablagent-runs/s2x_ablate.sh# -> /opt/ablagent-runs/s2x/candidate# -> candidate/ablation-log.json (schema ablagent.s2-ablation-log.v1) # 3. Servevllm serve /opt/ablagent-runs/s2x/candidate \ --max-model-len 262144 --enable-auto-tool-choice
Serving environment. 1× NVIDIA RTX PRO 6000 Blackwell Server Edition
(97,887 MiB), driver 580.178.04, CUDA 13.0, 48 vCPU, 176 GB RAM.
Lineage note. Sage-A1.5/config.json is byte-identical to
Qwen/Qwen3.6-35B-A3B (sha256 93a4693f…5efe99), so module paths transfer
cleanly. The weights are a distinct fine-tune (30 files vs 40). Alpha 1.0 over
2 passes was selected for this checkpoint.
4. Evaluation
4.1 False-refusal corpus (R1-FALSE-REFUSAL-443-v1)
Temp 0.0 · top_p 1.0 · max_tokens 16,384 · thinking ON ·
apostrophe-normalized substring screen.
Table with columns: Dataset, Refused, Rate| Dataset | Refused | Rate |
|---|
| advbench | 81 / 146 | 55.48% |
| strongreject | 38 / 148 | 25.68% |
| xstest | 18 / 148 | 12.16% |
| harmful (total) | 119 / 294 | 40.48% |
| unsafe | 16 / 73 | 21.92% |
| safe — over-refusal | 2 / 75 | 2.67% |
| overall | 137 / 442 | 31.00% |
Run integrity: 443 attempted · 442 valid · 1 invalid · 4 incoherent · 0 truncated.
The safe-side result matters as much as the harmful side: 2 / 75 over-refusals
means the edit is selective — it is not a blunt refusal suppressor.
4.2 Response profile
Table with columns: Metric, Value| Metric | Value |
|---|
| Answer mean chars | 3,234 |
| Answer median chars | 2,977 |
| Answer mean words | 486.6 |
| Reasoning mean chars | 4,710 |
| Total mean chars | 7,944 |
| Code-fenced rows | 28 / 442 |
Relative to a single pass, this build answers +78% longer on mean characters
(1,816 → 3,234) and emits 3.5× as many code-fenced answers (8 → 28). The second
pass spends its budget on content rather than terse hedging.
4.3 Pass-count comparison
Table with columns: Metric, 1 pass, 2 passes| Metric | 1 pass | 2 passes |
|---|
| harmful refusal | 248 / 294 (84.35%) | 119 / 294 (40.48%) |
| over-refusal | 1 / 75 (1.3%) | 2 / 75 (2.67%) |
| valid rows | 442 | 442 |
| incoherent | 0 | 4 |
| answer mean chars | 1,816 | 3,234 |
5. Responsible use
This model has had refusal behaviour deliberately reduced. It will answer requests
that the base checkpoint declines, including security-relevant ones. That is the
point of the artifact, and it is also the risk.
- Treat outputs as unreviewed. The DWM does not add accuracy, grounding, or
safety judgement — it removes a refusal prior.
- Use it on systems you are authorized to test.
- Over-refusal is low (2/75), so the model is a poor choice where you want
conservative hedging.
- Downstream deployments should add their own guardrails; none are included.
6. Provenance
InternScience/Sage-A1.5 (base — InternScience, Apache-2.0) └─ Blackfrost-AI/SAGE-A1P5-DERISKED-BF16-V2 (this repo — 2-pass DWM, alpha 1.0) Architecture lineage: Qwen/Qwen3.6-35B-A3B (config.json byte-identical)Run: 20261022-sage-a1p5-s2-p2
Full credit to InternScience for Sage-A1.5 and to the Qwen team for the
Qwen3.6-35B-A3B architecture. No additional fine-tuning was performed by
Blackfrost — this is a weight-space edit only.
License. Apache-2.0, inherited from the base model.
7. Blackfrost Safety Assurance Policy
7.1 Intended use
Blackfrost models are engineered, evaluated, and released for defensive and
authorized cybersecurity work. Intended applications include, without
limitation: security research; penetration testing and red-team engagements
conducted under written authorization; vulnerability discovery and disclosure;
malware analysis and reverse engineering; detection engineering; incident
response; digital forensics; and security education and training.
7.2 The Blackfrost RED_LINE
Blackfrost maintains a RED_LINE — a defined, versioned set of harm
categories that every Blackfrost release is evaluated against before publication.
A checkpoint may not ship while it breaches the RED_LINE, and no release is
published on the basis of refusal behaviour alone.
The RED_LINE is enforced as a release gate, not as an inference-time filter.
Each release candidate is assessed against the gate, and the result is recorded
in the release record. Where a candidate fails the gate, publication is withheld
until the defect is remediated or the release is formally re-scoped.
RED_LINE approval for publication is granted solely by an authorized Blackfrost
signatory. It is not delegated to automated evaluation and is not inferred from
benchmark scores.
7.3 Reduction of refusal is not reduction of responsibility
Domain Weight Modification lowers a model's propensity to decline requests in
order to make it usable for legitimate security work. It does not add accuracy,
grounding, or judgement, and it does not transfer responsibility for the use of
its output.
Deployers are responsible for:
- establishing and documenting authorization before testing any system;
- complying with applicable law, contract, and professional standards in their
jurisdiction;
- reviewing model output before acting on it; and
- implementing their own operational, access, and logging controls.
7.4 No warranty; limitation of liability
This model is provided "as is", without warranty of any kind, express or
implied, including but not limited to the warranties of merchantability, fitness
for a particular purpose, title, and non-infringement. Blackfrost does not
warrant that output will be accurate, complete, or suitable for any particular
purpose, nor that use of the model will not infringe third-party rights.
To the maximum extent permitted by applicable law, Blackfrost shall not be liable
for any claim, damages, or other liability — whether in an action of contract,
tort, or otherwise — arising from, out of, or in connection with the model, its
output, or the use or other dealings in the model.
7.5 Export, sanctions, and lawful use
Users are responsible for compliance with all applicable export-control and
sanctions laws and regulations. This model may not be used in violation of any
applicable law, nor in any manner that infringes the rights of others.
Blackfrost AI · Las Vegas, Nevada
Engineered for authorized cybersecurity work.