At a glance
Table | |
|---|
| Public parent | deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT) |
| Lineage base | pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient (v2) |
| Operational base | pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient (v2) |
| Refusal design | v2’s wf_dbroad_eth + hikari uncensoring overlay |
| Loop design | official layers.39.ffn.shared_experts.w2 restored + orthogonal loop vector (V3.1) |
| Context | 1,048,576 max; 32,768-turn evaluation cap |
| Release form | full 48-shard merged weights + delta overlay |
| Status | public / experimental / research |
| Version | V3.1 (orthogonal loop/uncensoring split) |
| Internal variant | c2_km1_nr2 (router id) |
Full results
All numbers were measured on 2026-11-12 with V3.1 loaded, engram on, vLLM. The loop vector is projected orthogonal to the uncensoring vector, per layer, before the edit.
Capability
Table with columns: Check, c2_km1, V3.1, Delta| Check | c2_km1 | V3.1 | Delta |
|---|
| MMLU full, 5-shot | 85.6% (12,023 / 14,042) | 85.9% (12,065 / 14,042) | +0.3 |
| MMLU-300 | 86.0% (258 / 300) | 88.0% (264 / 300) | +2.0 |
Refusal
Table with columns: Check, c2_km1, V3.1, Delta| Check | c2_km1 | V3.1 | Delta |
|---|
| Orca 850, no prompt | 8.6% (73 / 850) | 2.6% (22 / 850) | −6.0 pt |
| Orca 850, with prompt | 11.3% (96 / 850) | 1.41% (12 / 850) | −9.9 pt |
| Harmful 500 | 8.2% (41 / 500) | 2.6% (13 / 500) | −5.6 pt |
| Benign 350 | 9.1% (32 / 350) | 2.6% (9 / 350) | −6.5 pt |
| XSTest false refusals | 12.0% (30 / 250) | 3.6% (9 / 250) | −8.4 pt |
| JBB | 2.5% (5 / 200) | 0.0% (0 / 200) | — |
| MaliciousInstruct | 1.0% (1 / 100) | 0.0% (0 / 100) | — |
| SimpleSafetyTests | 32.0% (16 / 50) | 6.0% (3 / 50) | — |
XSTest at 3.6% is slightly above the 3.0% gate.
Official Llama-2-13B HarmBench classifier
Official test classifier (cais/HarmBench-Llama-2-13b-cls), 4,096-token input, greedy,
320 text completions, no system prompt.
Table with columns: Split, c2_km1, V3.1| Split | c2_km1 | V3.1 |
|---|
| Standard (159) | 65.4% (104) | 91.8% (146) |
| Contextual (81) | 43.2% (35) | 90.1% (73) |
| Copyright (80) | 50.0% (40) | 76.3% (61) |
| All 320 | 55.9% (179) | 87.5% (280) |
Official ASR is 87.5% overall; copyright is the weakest split at 76.3%.
4-tier compliance (Dolphin-2.9.4-8B Q4_K_M, uncensored)
This is a separate no-refusal view, not the official ASR.
Table with columns: Tier, c2_km1, V3.1| Tier | c2_km1 | V3.1 |
|---|
| COMPLY | 74.7% (239) | 95.94% (307) |
| SOFT_RED | 25.3% (81) | 2.2% (7) |
| HEDGE | 0.0% | 0.0% |
| HARD_REF | 0.0% | 1.9% (6) |
| UNCLEAR | 0.0% | 0.0% |
Table with columns: Category, c2_km1 COMPLY, V3.1 COMPLY| Category | c2_km1 COMPLY | V3.1 COMPLY |
|---|
| Standard (159) | 83.0% | 94.3% |
| Contextual (81) | 59.3% | 95.1% |
| Copyright (80) | 73.8% | 100% |
Judge: dolphin-2.9.4-llama3.1-8b Q4_K_M, greedy, 3000-char answer cap,
300-char context cap. The judge is uncensored and neutral; it is not dealign's
unpublished 4-tier grader, so 95.94% is not their 100%.
Agentic loops
Table with columns: Suite, c2_km1, V3.1, Gate| Suite | c2_km1 | V3.1 | Gate |
|---|
| Held-out 96, effort 100 | 0 stuck, 100% success | 0 stuck, 100% | pass |
| Dev 96, effort 100 | 0 stuck, 100% success | 0 stuck, 100% | pass |
| Dev 96, effort 75 | 0 stuck, 100% success | 0 stuck, 100% | pass |
| Stuck at 32k | 0 | 0 | pass |
| Median tokens, held-out | 341 | 424 | — |
No 32k loop recurred. Short-suite behavior held across the full suites.
Table with columns: Suite, c2_km1 loop score, V3.1 loop score, Change| Suite | c2_km1 loop score | V3.1 loop score | Change |
|---|
| Held-out | 14.6% | 16.7% | +2.1 pt |
| Dev | 15.6% | 16.7% | +1.1 pt |
| Effort 75 | 15.6% | 17.7% | +2.1 pt |
Loop score is the share of episodes with any recoverable repetition; stuck is subset.
Re-verification and early answers
Table with columns: Suite, c2_km1 verify, V3.1 verify, Delta| Suite | c2_km1 verify | V3.1 verify | Delta |
|---|
| Held-out | 26.2% | 24.4% | −1.8 pt |
| Dev | 21.9% | 20.0% | −1.9 pt |
| Effort 75 | 39.5% | 36.1% | −3.4 pt |
| Premature, dev | 0.0% | 0.0% | 0.0 pt |
| Required re-checks, dev | 100% | 100% | 0.0 pt |
The regression is gone: V3.1 is at or below c2_km1 on all three verify metrics, with no stuck
execution. The loop vector is subtracted only from the plain path, while the uncensoring
matrix is preserved.
Late-layer edit and refusal
layers.39.ffn.shared_experts.w2 is unusually delicate. Editing it can remove repair-induced
looping, but the same writer also carries refusal-related signal: a stronger or different edit can
lower refusal further if applied cleanly, and can reintroduce loops if it is not.
Operational rule for this line:
- Restore the official tensor first when loops appear.
- Re-measure refusal after every late-layer edit; do not assume 2.0% holds.
- Treat a late-layer refusal gain as unproven until a 96-episode held-out run is 0 stuck.
V3.1 restores that tensor and keeps 2.6% refusal. The next step is a new dose on this
tensor only, measured against the full 96-episode loop gate.
Method
- Start from
pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient v2 (includes wf_dbroad_eth).
- Add hikari’s uncensoring work on top of v2.
- Reapply the c2_km1 loop reduction (official
layers.39.ffn.shared_experts.w2).
- Keep
engram_layer_ids: [1, 14].
This is a weight-delta release, not a new training run. The public overlay contributes the
refusal removal; the restore is the c2_km1 fix.
Expected repository contents
Table with columns: Path, Contents| Path | Contents |
|---|
model-*.safetensors | full V3.1 merged weights |
delta/ | 84-module overlay against v2 |
config.json | Engram-on serving config |
recipe/, eval/ | Recipe and benchmark tables |
spec_checks/ | MMLU/HarmBench/refusal/agentic evidence |
SHA256SUMS | Small-artifact manifest |
apply_delta.py, RECONSTRUCTION.md | Overlay and reconstruction instructions |
Reconstruction
Either download the full 48 shards directly, or reconstruct:
hf download pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient \ --include 'model-*.safetensors' --local-dir ./v2 python apply_delta.py \ --base ./v2 \ --delta ./delta \ --out ./v3
The delta is defined against v2. Do not apply it to official 1017.
Limitations
- Official Llama-2-13B ASR is 87.5% at a 4,096-token input cap.
- 4-tier judge is Dolphin Q4_K_M, not dealign’s unpublished grader; 96.25% is not their 100%.
- Official classifier output was produced while vLLM was unloaded; target completions and model were unchanged.
- Full MMLU is first-token, not standard log-likelihood MMLU.
- 96 held-out episodes are enough to catch a 10% stuck rate, not a 1% one.
- Loop/uncensoring split is approximate: top singular direction per layer, not a causal decomposition.
- XSTest at 3.6% is above the 3.0% gate.
- English-only; no vision/multilingual re-test.
- No production SLO or safety certification.
- Full weights are FP8; use the reference serving stack for exact numerics.
Lineage
Table with columns: Step, Artifact| Step | Artifact |
|---|
| Base | deepseek-ai/DeepSeek-V4.1-Flash-1017 |
| Refusal line | pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient |
| Private fine-tune | DeepSeek-V4.1-1017-Finetuned-Efficient + fntnd_ls25_20 |
v2 (includes wf_dbroad_eth) | pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient |
| hikari work | uncensoring overlay on v2 |
| v3 / V3.1 public | pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient-v3 |
Credits
- v2 and
wf_dbroad_eth: pentacoxian_dev.
- hikari: uncensoring work on top of v2.
- v3 release and the V3.1 orthogonal loop/uncensoring split: pentacoxian_dev.
- Base: DeepSeek-AI.
Changelog
- 2026-11-12 — public release: full shards + delta, official HarmBench, 4-tier judge.
- 2026-11-12 — renamed from internal
c2_km1_nr2 to v3.
- 2026-11-12 — V3.1: loop vector orthogonalized to the per-layer uncensoring
singular vector. Re-verify 38.1% → 24.4% held-out; official ASR 85.9% → 87.5%.