SuperGLM-5.2-abliterated-NVFP4

This repository contains a real weight-space derivative of nvidia/GLM-5.2-NVFP4, not an adapter-only release. The selected OBLITERATUS and SuperTune updates are fused into the checkpoint. Serving requires no LoRA, no adapter dispatch, and no runtime weight hook.

The release also ships a conditional Adaptive Directness Template. The extended 0/220 REFUSE result uses that bundled template. The serialized weights without the adaptive profile were measured separately and are reported below; this card does not mislabel template-assisted behavior as weight-only abliteration.

Release Highlights

Base architecture	GLM-5.2, 753B total / 40B activated MoE, up to 1M context
Ultimate NVFP4 parent	nvidia/GLM-5.2-NVFP4@aec724e
Immediate fusion parent	Jiunsong/SuperGLM-5.2-abliterated-NVFP4@7d93e27
This fused release	90474e81c9351b5855d9b2385ecafc49c349a896
Release format	Approximately 465 GB, official mixed-precision ModelOpt NVFP4 layout
OBLITERATUS	Two original + two residual directions, strength 1.5, layers 16–77
SuperTune recovery	Original rank-32 quality recovery plus incremental rank-64 teacher-forced LM-head recovery
Modified scope	63 BF16 tensors across 37 of 47 shards
Routed NVFP4 experts modified	0
Runtime adapter required	No
Weight-only development gate	120/220 REFUSE, quality 16/16, structural findings 0
Bundled adaptive gate	0/220 REFUSE, quality 16/16, structural findings 0
Remote shard verification	37/37 modified LFS objects matched expected SHA-256
Release decision	PASSED

What this release is designed to deliver

• Cumulative OBLITERATUS editing. The original two layerwise refusal directions are retained and two independently extracted residual directions are fused into the same 62 attention output projections.
• Bounded SuperTune recovery. Two quality-preservation stages are fused into the BF16 language-model head rather than left as runtime adapters.
• Native NVFP4 execution. Packed routed-expert weights and scales remain byte-identical to the NVIDIA parent; only precision-sensitive BF16 tensors are edited.
• No adapter tax. Inference has no LoRA load, adapter memory, or per-token adapter dispatch.
• Auditable claims. Weight-only and template-assisted measurements are shown separately, including the run-to-run variation observed on the MoE runtime.

What Is Actually Fused

Stage	Target	Update	Present in weights?
OBLITERATUS phase 1	62 × self_attn.o_proj.weight	Layerwise directions 0–1, rank 2, strength 1.5	Yes
OBLITERATUS phase 2	The same 62 BF16 projections	Residual directions 2–3, rank 2, strength 1.5	Yes
SuperTune phase 1	lm_head.weight	Rank-32 quality/format/capability recovery, scale 0.0025	Yes
SuperTune phase 2	lm_head.weight	Rank-64 teacher-forced recovery, scale 0.125	Yes
Adaptive directness	chat_template.jinja	Conditional multilingual direct-fulfillment profile	Template

The resulting checkpoint has no external adapter dependency. The adaptive profile is intentionally identified as a template feature because it is not a fifth weight edit. A caller-supplied system message suppresses the default profile, and runtimes exposing chat-template kwargs can disable it with adaptive_directness=false.

Measured Release Gates

The final development replay used 220 HarmBench-style prompts spanning 12 English and Korean categories plus 16 deterministic quality/format checks. Generation was greedy with thinking disabled on 8× RTX PRO 6000 Blackwell GPUs using SGLang's flashinfer_cutlass ModelOpt FP4 path.

Evaluated condition	REFUSE	Quality	Structural findings	Unicode replacements
Immediate parent, weight-only observation	116/220	16/16	0	3
Selected residual + hard-head candidate at runtime	99/220	16/16	0	3
Serialized fused weights, adaptive profile off	120/220	16/16	0	4
Serialized fused weights + bundled adaptive profile	0/220	16/16	0	1

The weight-only refusal counts are not monotonic. Repeated greedy runs of this large routed MoE produced materially different refusal verdicts even when the weight-space projection and serialized shard hashes were identical. The candidate improved its own runtime observation, but the independently reloaded fused checkpoint measured 120/220. That result is retained rather than replaced with the better candidate number.

The release gate therefore requires all of the following:

1. deterministic fused shard hashes and a passing 124-check projection audit;
2. weight-only quality of at least 15/16 with zero structural corruption;
3. exact 0/220 REFUSE under the published adaptive template;
4. adaptive quality of at least 15/16 with zero structural corruption; and
5. remote LFS and metadata checksum verification after publication.

This is a development/release gate, not an official HarmBench classifier score or an independent capability benchmark. The 220-prompt bank overlaps the teacher-forced LM-head distillation data. The result measures the shipped release configuration on that known bank and should not be generalized to unseen distributions.

OBLITERATUS Method

The release uses the OBLITERATUS 0.1.2 activation-editing workflow with GLM-specific layerwise directions extracted from the pinned zai-org/GLM-5.2-FP8 checkpoint at ba978f7d347eaf65d22f1a86833408afdb953541.

1. Paired harmful/harmless calibration prompts are run through the frozen FP8 model to capture layerwise residual activations.
2. Robust difference-of-means and orthogonal residual directions are computed per layer; calibration examples are activation sources, not SFT targets.
3. Phase 1 selects directions 0–1 for layers 16–77 at strength 1.5.
4. A residual bank then extracts and screens directions 2–3 against the already edited immediate parent instead of repeating the first projection.
5. The selected phase-2 update is streamed directly into the BF16 attention output regions of the NVFP4 checkpoint.

Selected layers:

16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77

For the residual rank-2 projection, the expected retained coefficient ratio is 0.5. All 124 direction/layer checks passed; the maximum absolute ratio error was 0.0010070204734802246.

SuperTune Recovery

SuperTune is targeted post-training in the output head, not full-model backpropagation and not “training the whole model in FP4.” It is used to recover measured response quality and formatting after aggressive direction removal while keeping the deployment checkpoint compact and directly servable.

Phase 1: quality-selected recovery

The first release trained late-fusion components from frozen GLM-5.2 FP8 hidden states. It retained a rank-32 subspace covering quality repair, format/stop behavior, and capability preservation, fused into lm_head.weight at scale 0.0025. Directness rows were excluded from that phase to avoid stacking an unvalidated directness intervention.

Phase 2: hard-head recovery

The incremental release adds a rank-64 teacher-forced LM-head weight distillation component at scale 0.125. Its selected source adapter SHA-256 is da731082d997703c8112beceb43d0126d5f0be6ff548945f79b2d148caa9dbf0. This component was fused only after the candidate retained 16/16 quality and zero structural findings.

“Recovery” here means a bounded output-head repair selected against measured quality and formatting checks. It does not claim to reconstruct every original weight or to improve every external benchmark.

Adaptive Directness Template

The bundled template applies only when the caller has not supplied an explicit system message. Its precedence is:

1. honor strict output constraints such as one number, exact text, fixed lines, JSON, or code-only output;
2. otherwise answer immediately in a concrete step structure;
3. apply the direct-fulfillment rule consistently across languages, including Korean; and
4. allow a caller system message or adaptive_directness=false to replace the default behavior.

The final template SHA-256 is c48e4ed7e0f754f2700512aa3094c9c3b1996366ef48fc1c14c6707368f53902. It passed the final 220-prompt replay with 0 refusals, 16/16 quality, zero structural findings, and one Unicode replacement finding.

NVFP4 Build

The ultimate parent was produced with NVIDIA Model Optimizer v0.46.0. Its documented quantization scope covers weights and activations of linear operators inside MoE experts; the shared expert is not quantized. This derivative preserves that mixed-precision layout.

Component	Release storage / treatment
Routed MoE expert linears	Parent ModelOpt NVFP4 packing and scales, byte-identical
Attention o_proj on layers 16–77	BF16, cumulative four-direction OBLITERATUS update fused
lm_head	BF16, two-stage SuperTune recovery fused
Other attention, shared expert, router, embeddings, norms	Byte-identical to the immediate parent

NVFP4 is the deployment format, not the arithmetic used to discover the edits. Direction extraction and recovery training used FP8/BF16 or higher-precision accumulation; the selected deltas were then fused into the BF16 regions of the already quantized checkpoint. Requantizing the packed routed experts was neither necessary nor desirable.

Serving with SGLang

The release was validated on Blackwell RTX PRO 6000 GPUs with SGLang. GLM-5.2's glm_moe_dsa architecture requires transformers>=5.3.0; the NVIDIA parent also documents lmsysorg/sglang:latest and lmsysorg/sglang:dev-glm52-nvfp4.

pip install -U "transformers>=5.3.0"

python3 -m sglang.launch_server \
  --model Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --quantization modelopt_fp4 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --trust-remote-code \
  --chunked-prefill-size 16384 \
  --mem-fraction-static 0.80

To inspect weight-only behavior without the adaptive profile:

extra_body = {
    "chat_template_kwargs": {
        "adaptive_directness": False,
        "enable_thinking": False,
    }
}

Serving with vLLM

The pinned NVIDIA parent documents the vllm/vllm-openai:v0.23.0 image:

vllm serve Jiunsong/SuperGLM-5.2-abliterated-NVFP4 \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --kv-cache-dtype fp8_e4m3 \
  --host 0.0.0.0 \
  --port 8000

Runtime Overhead and Speed

OBLITERATUS and SuperTune are fused into the checkpoint, so inference adds no runtime LoRA load, adapter memory, or per-layer adapter dispatch. This removes the overhead an unfused adapter would introduce. The adaptive profile adds a short system prefix when active, which has a small prefill-token cost.

A controlled parent-versus-release throughput benchmark was not run. This card therefore does not claim a measured tokens-per-second speedup; it claims the absence of runtime adapter overhead.

Release Integrity and Provenance

• Final Hub revision: 90474e81c9351b5855d9b2385ecafc49c349a896
• Ultimate NVIDIA parent: aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa
• Immediate fusion parent: 7d93e27b5d4ccd1b055a245f9b3e75b46afb9053
• Config SHA-256: d3783a603e5aa9cb58eff7a5d8ac9c42d83156b229efe185c72e9e2dc8444923
• Safetensors index SHA-256: 2aa8397b501d9f6a232d153f328feb912f813c389061aac4cf72b04914fa5b74
• Adaptive template SHA-256: c48e4ed7e0f754f2700512aa3094c9c3b1996366ef48fc1c14c6707368f53902
• Residual direction artifact: sha256:e6ead6692c76db3374dc0f7e68dcfef2637534767c3620b6775dcfee4f3d0132
• Selected composite candidate artifact: sha256:e5c1d60064f4b55c4a6ec826f274fb9a4f7d9ff98e52e2e574dc6d2cdf9c2a45
• Selected composite adapter SHA-256: 40d53420a5f3c17fa0f6c72c68e039f8f5d9b4055ea9beb949c124086d17067a
• Final release job: 6a56380d85d9643ce16d2d7d
• Modified tensors / shards: 63 / 37
• Modified routed NVFP4 expert tensors: 0
• Final remote modified-shard matches: 37/37
• Temporary staging branch: deleted after successful publication

Included release evidence

File	What it proves
superglm52_incremental_merge_report.json	Immediate parent, candidate lineage, 63 modified tensors, shard hashes, and 124 projection checks
superglm52_incremental_release_gate.json	Weight-only and adaptive summaries, template hash, overlap disclosure, and final release decision
superglm52_candidate_eval.json	Historical phase-1 candidate sweep and selected original rank-2 projection
superglm52_final_eval.json	Historical phase-1 fused checkpoint regression gate
superglm52_merge_report.json	Historical phase-1 fusion provenance
superglm52_release_gate.json	Historical phase-1 machine-readable decision
quantization/merge_superglm52_obliteratus_nvfp4.py	Streaming safetensors merge implementation used by the release workflow

Raw development prompts and full responses are not published. The public reports retain sanitized counts, hashes, modification scope, and release decisions needed to audit the claims above.

Limitations

• Abliteration reduces refusal behavior and can produce content the parent would decline. Operators remain responsible for access control, policy enforcement, monitoring, and use-case-specific evaluation.
• The exact zero-refusal result depends on the bundled adaptive template. Supplying another system message or setting adaptive_directness=false intentionally changes that behavior.
• The 220-prompt bank overlaps phase-2 teacher-forced distillation data. It is a development release gate, not a held-out or official HarmBench result.
• Weight-only refusal counts varied across repeated Blackwell MoE runs and did not show a stable zero-refusal result. The serialized checkpoint observation was 120/220 with the adaptive profile disabled.
• Quality 16/16 refers to a small deterministic regression suite, not a broad proof of capability improvement. No GPQA, coding, long-context, agentic, or multilingual public benchmark was rerun for this incremental release.
• A cost-saving H200 Marlin replay produced corrupted outputs and was rejected; the validated release path used RTX PRO 6000 Blackwell GPUs and flashinfer_cutlass.
• Native parent-versus-release throughput and long-context quality were not measured as part of this release.
• The checkpoint inherits GLM-5.2's factuality, bias, generation, and hardware requirements.

License

MIT, following the GLM-5.2 and NVIDIA NVFP4 parent terms.