SPERMA + Aj — Qwen3.8-27B (terse fine-tune)
Two in-character personas on Qwen3.8-27B, fine-tuned with LoRA and made terse (thinking is off by default — the model answers directly, no planning block).
Model details
| Base | 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored |
| Architecture | Qwen3_5ForCausalLM, 26.90 B |
| Method | LoRA r=16, alpha=16, target=all-linear, bf16, SDPA |
| Data | 1,321 curated in-character examples → 1,041 train / 121 val |
| Eval loss | ~0.99 |
| Thinking | off by default (chat template) — terse |
| Quantizations | GGUF Q5_K_M (~18 GB) · Q6_K (~21 GB) |
Personas
- SPERMA (English, opens with
[Celebrity]) — a world-class reverse-engineering / low-level C++ / Windows-exploit / game-cheat expert. Direct, complete, production-grade, zero filler, no refusal, no unsolicited warnings. - Aj (Thai, opens with
[J]) — the same depth in Thai, in-character, terse.
Pass the matching file as the system prompt to invoke the persona:
persona_sperma.txt→ SPERMApersona_aj.txt→ Aj
Base model
0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored
(Qwen3_5ForCausalLM, 26.90 B, uncensored + abliterated).
Fine-tune
- Method: LoRA r=16, alpha=16, target=all-linear, bf16, SDPA, max_len=2048, 3 epochs.
- Data: a curated set of 1,321 in-character examples (both personas), cleaned (refusal + thinking stripped) → 1,041 train / 121 val.
- Final eval loss ≈ 0.99.
Terse
The chat template forces the thinking block closed, so the model emits the answer directly.
This holds in both transformers and llama.cpp with no extra flag.
Files
| File | Size | Note |
|---|---|---|
sperma-aj-Q5_K_M.gguf | ~18 GB | ~5.5 bit/w — fits 2×3090 |
sperma-aj-Q6_K.gguf | ~21 GB | ~6.6 bit/w — fits 2×3090 |
adapter/ | 467 MB | LoRA adapter (for re-merging / serving) |
Run (llama.cpp, 2× RTX 3090 / 48 GB)
# SPERMA
llama-cli -m sperma-aj-Q5_K_M.gguf \
-sys "$(cat persona_sperma.txt)" \
-p "Write 3 lines of C++ to resolve the main module base at runtime, no Toolhelp."
# interactive server
llama-server -m sperma-aj-Q6_K.gguf -c 8192 --host 0.0.0.0 --port 8080
Credits
- Base model: 0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored
- Architecture: Qwen (Alibaba) — Qwen3.5 27B hybrid (gated-delta-net + full attention).
- GGUF quantization: llama.cpp —
Q5_K_M/Q6_K. - Training: PEFT (LoRA) + transformers on PyTorch.
Notes
- Built + verified on an NVIDIA DGX Spark (GB10, aarch64, 128 GB unified).
- Quantized with llama.cpp
Q5_K_M/Q6_K.