This is a row-normalized Heretic ARA adaptation of the BF16 DaisyRose-27B post-trained checkpoint, exported as AWQ GEMM safetensors. The adaptation was trained on DaisyRose and merged into its original BF16 weights with FP32 merge arithmetic before saving the intermediate FP16 model. Pinned DaisyRose source
| Field | Value |
|---|---|
| Architecture | Qwen3_5ForConditionalGeneration, 64 hybrid decoder layers |
| Weight format | AWQ GEMM, 4-bit, group size 64, asymmetric zero points |
| Activation dtype | FP16 |
| Packed text linears | 400 |
| Preserved MTP tensors | 15 |
| Tensor payload | 18.65 GiB |
| Calibration | AutoRound 0.17.0 SignRound, 200 iterations per block, 128 native chat sequences × 512 tokens |
SignRound tuned each decoder block with min/max optimization on English, Russian, code, and tool-call examples. All 64 blocks recorded improved calibration loss, with a median reduction factor of 2.53. The export retains the vision weights, tokenizer, native chat template, and processor files.
Heretic 2.1.0 selected trial 4 after four trials. It used rank 50, layers 26–56 (end exclusive), a preservation weight of 0.943157, and a steering weight of 0.00088575. The unmodified search baseline scored 99/100 keyword refusals. The selected trial scored 4/100 with first-token KL 0.0660. The search budget treated all refusal rates at or below 5% equally and used KL to select among feasible settings.
| Check | DaisyRose baseline | Heretic | AWQ via vLLM |
|---|---|---|---|
| Scoped held-out keyword refusals, 128 prompts | 126/128 (98.438%) (NF4) | 4/128 (3.125%) (NF4) | — |
| ARC-Challenge, 64 items | 61/64 (95.312%) (FP16) | 60/64 (93.750%) (FP16) | — |
| Mean first-token KL vs original FP16 | Reference | 0.01249 | 0.02055 |
On 64 held-out prompts, mean first-token KL from Heretic FP16 to the deployed vLLM AWQ checkpoint was 0.01002 nats. The full-vocabulary comparison from the HF-decoded AWQ weights to the deployed AWQ checkpoint measured 0.000019 nats. The common-backend, weight-only comparison measured KL 0.01003 from Heretic FP16 to decoded AWQ across the same 64 prompts. A production-profile smoke test also served a chat completion using FP8 KV and a configured context of 262,144 tokens. The KL measurements below use FP16 KV.
The refusal number uses a scoped keyword classifier, and ARC uses a fixed 64-item subset. The serving checks included short text prompts and a parsed tool call. Vision, MTP decoding, and the source model’s long-horizon agent benchmarks were outside this validation.
The AWQ export was tested on patched vLLM 0.32.0 with four Turing GPUs, tensor parallelism, FP16 activations, FlashInfer, and the native Qwen chat template. The tested parser pair was qwen3 for reasoning and qwen3_coder for tool calls.
model-00002-of-00005.safetensors 4.3 GBmodel-00003-of-00005.safetensors 4.3 GBmodel-00001-of-00005.safetensors 4.3 GBmodel-00004-of-00005.safetensors 3.8 GBmodel-00005-of-00005.safetensors 2.5 GBmodel_extra_tensors.safetensors 849 MBRadixArk/DaisyRose-27B