license: apache-2.0
base_model:
- YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4
base_model_relation: quantized
library_name: llama.cpp
tags:
- gguf
- llama.cpp
- quantized
- YFC-112358
- Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4
Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4-GGUF
GGUF quantizations of YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4.
Quantizations
- BF16 (50.9 GB)
- Q8_0 (27.0 GB)
- Q6_K (20.8 GB)
- Q5_K_M (18.1 GB)
- Q4_K_M (15.6 GB)
- IQ4_NL (15.0 GB)
Converted with llama.cpp.
Source Model
YFC-112358/Qwen3.8-27B-Della-Deckard-Fable-Qwopus-ColdFusion-v4
Chinese to English Translation of Source Model Card
Overview of the Three‑Stage Merge
Stage 1: Use DELLA to fuse three different‑origin models into a single “general‑intelligence composite G.” Stage 2: Linearly stack the incremental changes from Qwopus3.6‑27B‑Fusion onto G, without modification. Both stages use the true ancestor Qwen/Qwen3.6‑27B as the base model (not any descendant), so all task vectors Δ exist in the same reference frame — making linear addition meaningful. Stage 3: Move the entire result from Stage 2 onto Cold‑Fusion‑GAIN‑V1.1 (Qwen3.8): out = Cold‑Fusion + (stage2 − Qwen3.6) This means: take the task vector relative to Qwen3.6 and transplant it onto the Qwen3.8‑based Cold‑Fusion model. This step is precise, no pruning, no randomness — but cross‑generation, so its validity depends on whether the two generations’ weight bases align.
Version 3 Coefficient Note
This v3 keeps W_V3 = 0.30 (not 1.00):
output = Qwen3.6 + 0.30·T(Δ_v3) + 1.00·Δ_ColdFusion
Because β = 1.00, the base is fully Cold‑Fusion (Qwen3.8).
α controls how much of the 3.6‑side task vector is added.
α = 0.30 is empirically calibrated, not arbitrary:
Cold‑Fusion’s displacement relative to official Qwen3.8 is only ≈0.7%‖W‖
(i.e., 98.7% generational difference, 1.3% its own fine‑tuning)
Δ_v3 is ≈13%‖W‖ on MLP and nearly orthogonal (cos ≈ +0.005)
If α = 1.00, Δ_v3 would drown out Cold‑Fusion’s behavior by ~15×.
α = 0.30 keeps MLP displacement at ~4.6× Cold‑Fusion’s own shift.
Companion Adapter
A matching LoRA adapter exists: YFC‑112358/Qwen3.8‑27B‑Della‑Deckard‑Fable‑Qwopus‑ColdFusion‑v4‑LoRA Stage‑3’s T(Δ_v3) is saved as a rank‑48 PEFT LoRA α becomes a knob at load time: lora_alpha = α × 48 Can be applied to any Qwen3.8‑series model embed_tokens and lm_head are not included due to memory limits
Implementation Notes
Stage 1’s G is not saved to disk; computed in‑memory and immediately used for Stage 2. Mathematically identical to saving/loading, but avoids ~54GB of I/O. ε automatically narrowed to legal limits to satisfy mergekit density constraints. Stage 2 uses pure linear addition, not DELLA. Because Qwopus’s increments are tiny (0.2–0.8% in MLP), pruning would remove signal and keep noise. Stage 3 is the largest uncertainty: cross‑generation linear transplant. It only works if Qwen3.8 is a continuation of Qwen3.6. Cold‑amp measured 0.1217, within the typical same‑base fine‑tuning range (0.02–0.20). If cold‑amp were near 1.0, treat this model as experimental.
Fusion Health Metrics (Measured During Construction)
Metric Meaning Median amp total increment / ancestor weight 0.0202 share G’s contribution relative to Qwopus 26.002 kill₂ % of non‑zero elements removed by Stage‑2 sign election 0.0% cold_amp ‖Qwen3.8·ColdFusion − Qwen3.6‖ / ‖Qwen3.6‖ 0.1217 Notes: Qwopus’s increments are naturally tiny, so share is large. kill₂ = 0% because Stage 2 uses pure linear addition (no DELLA pruning).
Sampling Recommendations
Start with: temperature = 0.7 top_p = 0.8 top_k = 20 Tokenizer/config come from Cold‑Fusion (Qwen3.8), since weights are anchored there.
Deviation From Original Recipe
Stage 2 is strict linear addition, not DELLA. Stage 3 is cross‑generation linear anchoring, the main uncertainty. No TIES sign election — because v2 and Fable vectors are nearly orthogonal; election would erase ~13% of v2’s unique features.