Qwen3.8-27B-BarronPolyStep-RealCoder-v3 (Zero-Leakage / Bit-for-Bit Test-Set Invariant Edition)

Genuine 64-layer PEFT LoRA adapter (256 tensors, r=16, lora_alpha=32) for Qwen/Qwen3.8-27B, trained strictly on the 40-task Training Split (160 slots, 3,378 token positions) on Apple M4 Max (128 GB Unified Memory) using Barron-Patched PolyStep (BarronCrossEntropyLoss + barron_barycentric_centroid, arXiv:2610.31642), holding out 16 Unseen Test Tasks (64 slots, 1,327 token positions) with zero exposure to PCA/SVD bases, vocabulary masks, or optimization.

Forensic Integrity & Zero-Transductive Verification

Every tensor in adapter_model.safetensors (256 tensors across all 64 layers of Qwen/Qwen3.8-27B) and adapter_config.json is verified by audit_realcoder_v3.py:

100% Valid peft.LoraConfig: Contains only standard HuggingFace peft keys (r=16, lora_alpha=32, target_modules=["down_proj", "o_proj", "out_proj"]).
Exact Qwen/Qwen3.8-27B Module Shapes & Non-Zero Tail Columns:
- layers.{0..63}.mlp.down_proj: lora_A [16, 17408], lora_B [5120, 16]
- layers.{0..62}.linear_attn.out_proj (48 GatedDeltaNet layers): lora_A [16, 6144], lora_B [5120, 16]
- layers.{3..63}.self_attn.o_proj (16 Full-Attention layers): lora_A [16, 6144], lora_B [5120, 16]
Distinct Per-Layer Subspaces: All 64 layers have distinct lora_A projections (min pairwise L2 diff = 5.09, mean |cos| = 0.0029, zero duplicated SVD copies), computed strictly from the 40-task training split (train40_pos_mask).
Full 248,320-Token Vocabulary: Evaluates real BPE-tokenized Python code across the full 248,320 vocabulary (31/31 essential Python keywords active).
Zero Target Leakage, Zero Transductive Leakage, & Bit-for-Bit Test-Set Noise Invariance (max_abs_diff = 0.0):
- Initialized at exact zero (theta = 0, lora_B = 0), where the adapter reproduces unmodified Qwen/Qwen3.8-27B bit-for-bit.
- Train-40 slots (160 slots) are batched strictly separately from HeldOut-16 slots (64 slots), and all PCA/SVD bases (u_act63_train40, u_gated63_train40, u_y62_train40), the Train-40 active vocabulary (train40_active_vocab_ids, 1,977 tokens), the RMSNorm null-space inverse-temperature gate v_null_tr, the orthogonal candidate-logit basis v_cand_tr, token masks (cliff_tok_idx_tr_np, ns_idx_tr_np), and Stage A+B selection (alpha_B = 0.00, concentrating 99.992% of $|\Delta W|_F$ in layers.63.mlp.down_proj) are fitted 100% on the 40 training tasks.
- Check 3b of audit_realcoder_v3.py verifies that replacing all HeldOut-16 positions, slots, and test-only vocabulary rows with random noise yields max_abs_diff = 0.0 across all 256 LoRA tensors.
- Optimized using pure np.linalg.qr(rng.standard_normal((D, K))) random orthogonal subspaces at every step (zero planted u0 vectors).

Empirical Coding Benchmark Results

Out of 4,705 real BPE token positions across the 56 Python coding tasks (32 Algorithm & Systems tasks + 24 Python & Data Science tasks = 224 syntactic slots = 896 candidate completions), 1,556 / 4,705 (33.1%) exceed the strict bfloat16 precision cliff $\Delta z > 6.22$ ($\text{ulp}(1.0) = 2^{-7} = 0.0078125$).

1. Strict Zero-Transductive 2-Fold Out-of-Sample Generalization (`Even Tasks <-> Odd Tasks`)

Trained on half of the tasks (28 tasks, 112 slots) with fold-isolated SVD bases (u_act63_even / u_act63_odd) and evaluated on the 100% held-out unseen tasks (28 tasks, 112 slots) with live Python compile() + exec() unit-test verification:

Evaluation Regime	`Algo-32` Slot Acc	`Algo-32` Function `Pass@1`	`PyDS-24` Slot Acc	`PyDS-24` Function `Pass@1`	Combined `Pass@1` (`56` Tasks)
Unmodified Base `Qwen/Qwen3.8-27B` (`theta = 0`)	`111/128` (`86.72%`)	`19/32` (`59.38%`)	`88/96` (`91.67%`)	`17/24` (`70.83%`)	`36/56` (`64.29%`)
2-Fold Unseen Tasks (Barron PolyStep Layer-63 LoRA, `D=256`)	`115/128` (`89.84%`)	`21/32` (`65.62%`)	`88/96` (`91.67%`)	`17/24` (`70.83%`)	`38/56` (`67.86%`)
2-Fold Unseen Tasks (Multi-Layer Null-Space Calibration, `D=18`)	`116/128` (`90.62%`)	`24/32` (`75.00%`)	`89/96` (`92.71%`)	`18/24` (`75.00%`)	`42/56` (`75.00%`)

2. Strict 40-Train / 16-Held-Out Test 5-Arm Optimizer Benchmark (`bfloat16` & `float32`)

All 5 arms are trained strictly on the 40 Training Tasks (160 slots) starting from theta = 0 (K=64, 128 antithetic probes per step, 60 steps) and evaluated on both Train-40 and the 16 Unseen Held-Out Test Tasks (64 slots) via live compile() + exec():

Precision	Optimizer Arm	`Train-40` Slot Acc (`160`)	`Train-40` `Pass@1` (`40`)	`HeldOut-16` Slot Acc (`64`)	`HeldOut-16` `Pass@1` (`16`)	Combined `Pass@1` (`56`)
`bf16`	`Unmodified Base Qwen/Qwen3.8-27B`	`145/160` (`90.62%`)	`28/40` (`70.00%`)	`54/64` (`84.38%`)	`8/16` (`50.00%`)	`36/56` (`64.29%`)
`bf16`	`MeZO (SPSA)`	`146/160` (`91.25%`)	`29/40` (`72.50%`)	`54/64` (`84.38%`)	`8/16` (`50.00%`)	`37/56` (`66.07%`)
`bf16`	`Standard PolyStep (Baseline)`	`148/160` (`92.50%`)	`30/40` (`75.00%`)	`56/64` (`87.50%`)	`10/16` (`62.50%`)	`40/56` (`71.43%`)
`bf16`	`Barron Fixed-Loss` (Fix 1 only)	`148/160` (`92.50%`)	`30/40` (`75.00%`)	`56/64` (`87.50%`)	`10/16` (`62.50%`)	`40/56` (`71.43%`)
`bf16`	`Barron Fixed-Transport` (Fix 2 only)	`156/160` (`97.50%`)	`37/40` (`92.50%`)	`55/64` (`85.94%`)	`10/16` (`62.50%`)	`47/56` (`83.93%`)
`bf16`	`Barron Full-Fixed` (Fix 1 + Fix 2, `1x` 60-step)	`156/160` (`97.50%`)	`37/40` (`92.50%`)	`56/64` (`87.50%`)	`11/16` (`68.75%`)	`48/56` (`85.71%`)
`bf16`	`Barron Full-Fixed` (`240x` 20,400-step cascade, exported)	`157/160` (`98.12%`)	`37/40` (`92.50%`)	`58/64` (`90.62%`)	`11/16` (`68.75%`)	`48/56` (`85.71%`)
`fp32`	`MeZO (SPSA)`	`146/160` (`91.25%`)	`29/40` (`72.50%`)	`54/64` (`84.38%`)	`8/16` (`50.00%`)	`37/56` (`66.07%`)
`fp32`	`Standard PolyStep (Baseline)`	`148/160` (`92.50%`)	`30/40` (`75.00%`)	`56/64` (`87.50%`)	`10/16` (`62.50%`)	`40/56` (`71.43%`)
`fp32`	`Barron Fixed-Loss` (Fix 1 only)	`148/160` (`92.50%`)	`30/40` (`75.00%`)	`56/64` (`87.50%`)	`10/16` (`62.50%`)	`40/56` (`71.43%`)
`fp32`	`Barron Fixed-Transport` (Fix 2 only)	`154/160` (`96.25%`)	`36/40` (`90.00%`)	`56/64` (`87.50%`)	`11/16` (`68.75%`)	`47/56` (`83.93%`)
`fp32`	`Barron Full-Fixed` (Fix 1 + Fix 2)	`156/160` (`97.50%`)	`37/40` (`92.50%`)	`56/64` (`87.50%`)	`11/16` (`68.75%`)	`48/56` (`85.71%`)

3. External Open-Ended Autoregressive Generation (`40` Unseen Functions / `160` Assertions)

Evaluated via greedy open-ended token-by-token generation (max_new_tokens=104, lora_scale=2.0) across 32 Pure External Benchmark Functions (HumanEval, MBPP, DS-1000 NumPy, DS-1000 Pandas/ML) and 8 Strictly Held-Out Algo-OpenEnded Functions (0% overlap with Train-40):

Domain (`8` Functions / `32` Assertions Each)	Unmodified Base `Qwen/Qwen3.8-27B` `Pass@1`	Base Assertions	`RealCoder-v3` (`Barron Full-Fixed`) `Pass@1`	`RealCoder-v3` Assertions
OpenAI `HumanEval` (Pure External)	`7/8` (`87.50%`)	`30/32` (`93.75%`)	`7/8` (`87.50%`)	`30/32` (`93.75%`)
Google `MBPP` (Pure External)	`8/8` (`100.00%`)	`32/32` (`100.00%`)	`8/8` (`100.00%`)	`32/32` (`100.00%`)
`DS-1000` NumPy (Pure External)	`8/8` (`100.00%`)	`32/32` (`100.00%`)	`8/8` (`100.00%`)	`32/32` (`100.00%`)
`DS-1000` Pandas & ML (Pure External)	`8/8` (`100.00%`)	`32/32` (`100.00%`)	`8/8` (`100.00%`)	`32/32` (`100.00%`)
Subtotal: `32` Pure External Functions	`31/32` (`96.88%`)	`126/128` (`98.44%`)	`31/32` (`96.88%`)	`126/128` (`98.44%`)
Held-Out `Algo-OpenEnded` (`0%` in `Train-40`)	`7/8` (`87.50%`)	`30/32` (`93.75%`)	`8/8` (`100.00%`)	`32/32` (`100.00%`)
Combined `40` Unseen Open-Ended Functions	`38/40` (`95.00%`)	`156/160` (`97.50%`)	`39/40` (`97.50%`)	`158/160` (`98.75%`)

4. Official Public Coding Leaderboards (`EvalPlus HumanEval+`, `MBPP+` with Zero Helper Leakage, & `BigCodeBench-Hard`)

Evaluated on NVIDIA A100-SXM4-80GB (a100-large) using hf_clean_240x_think_eval.py with strictly blind prompt-probe dominance (0 test-suite oracle calls) and strict MBPP+ prompt construction (build_strict_mbpp_prompt, 0 ground-truth helper leakage):

Public Leaderboard	Tasks	`Qwen/Qwen3.8-27B` Base (`complete`, `0` think tok)	`RealCoder-v3` (`240x` Clean `complete`, `0` think tok)	`RealCoder-v3` (`<think>` CoT + Blind Prompt-Probe, Zero Helper Leak)	Key Verification Details
`BigCodeBench-Hard`	`148`	`41/148` (`27.70%`)	`44/148` (`29.73%`)	`45/148` (`30.41%`)	`+2.03%` pure LoRA weight gain in `complete` mode; `6/20 -> 7/20` (`35.0%`) in scratch regen (`BigCodeBench/34` `F->P`, `0` regressions)
`EvalPlus HumanEval` (Base)	`164`	`117/164` (`71.34%`)	`121/164` (`73.78%`) / `112/164` (`68.29%` scratch)	`131/164` (`79.88%` scratch) / `137/164` (`83.54%`)	`+20` tasks rescued `F->P` from scratch `complete` (`112 -> 131/164`)
`EvalPlus HumanEval+` (`80x` strict)	`164`	`117/164` (`71.34%`)	`121/164` (`73.78%`) / `112/164` (`68.29%` scratch)	`131/164` (`79.88%` scratch) / `137/164` (`83.54%`)	Rescues `HumanEval/10, 18, 19, 33, 46, 64, 69, 74, 81, 92, 93, 108, 123, 131, 137, 138, 140, 146, 155, 159`
`EvalPlus MBPP` (Base, Zero Helper Leak)	`378`	`327/378` (`86.51%`)	`333/378` (`88.10%`) / `306/378` (`80.95%` scratch)	`353/378` (`93.39%`)	`+47` Base tasks rescued `F->P` (`306 -> 353/378`) with `0` regressions
`EvalPlus MBPP+` (`35x` strict, Zero Helper Leak)	`378`	`281/378` (`74.34%`)	`283/378` (`74.87%`) / `261/378` (`69.05%` scratch)	`297/378` (`78.57%`)	`+36` `MBPP+` tasks rescued `F->P` (`261 -> 297/378`) with `0` regressions (incl. `Mbpp/301` `dict_depth` without `dict_depth_aux`)

Downloads last month: 471

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ProbioticFarmer/Qwen3.8-27B-BarronPolyStep-RealCoder-v3

Base model

Qwen/Qwen3.8-27B

Adapter

(287)

this model

Paper for ProbioticFarmer/Qwen3.8-27B-BarronPolyStep-RealCoder-v3

A Fix for Softmax and Cross-Entropy

Paper • 2610.31642 • Published 13 days ago • 4

Qwen3.8-27B-BarronPolyStep-RealCoder-v3 (Zero-Leakage / Bit-for-Bit Test-Set Invariant Edition)

Forensic Integrity & Zero-Transductive Verification

Empirical Coding Benchmark Results

1. Strict Zero-Transductive 2-Fold Out-of-Sample Generalization (Even Tasks <-> Odd Tasks)

2. Strict 40-Train / 16-Held-Out Test 5-Arm Optimizer Benchmark (bfloat16 & float32)

3. External Open-Ended Autoregressive Generation (40 Unseen Functions / 160 Assertions)