palimpsest-lora-adapters

Palimpsest — Literary Style LoRA Adapters

Mix multiple authors' writing styles in real-time using lightweight LoRA adapters on a single base model.

Each adapter captures one author's distinctive voice (~435MB). At inference time, adapters are weighted and blended per request — 70% Bulgakov + 30% Poe produces lyrical Gothic satire. No model reload, no merge, instant style change.

Available Adapters (Public Domain Authors)

Base Models

• Russian adapters: IlyaGusev/saiga_nemo_12b — Mistral Nemo 12B, Russian-tuned, abliterated
• English adapters: natong19/Mistral-Nemo-Instruct-2407-abliterated

Both base models must be converted to GGUF Q6_K format for use with llama.cpp.

How to Use

With llama.cpp (recommended)

Russian and English adapters use different base models and must run on separate servers:

# Russian server (Saiga Nemo 12B base)
llama-server \
  -m saiga-nemo12b-base-q6k.gguf \
  --lora bulgakov.lora.gguf \
  --lora-init-without-apply --no-jinja \
  --port 8080 --ctx-size 4096

# English server (Mistral Nemo abliterated base)
llama-server \
  -m nemo-abliterated-base-q6k.gguf \
  --lora lovecraft_en.lora.gguf \
  --lora doyle_en.lora.gguf \
  --lora poe_en.lora.gguf \
  --lora wilde_en.lora.gguf \
  --lora london_en.lora.gguf \
  --lora-init-without-apply --no-jinja \
  --port 8081 --ctx-size 4096

# Per-request mixing (within same language)
curl http://localhost:8081/v1/chat/completions \
  -d '{"messages": [{"role": "user", "content": "Write a scene..."}],
       "lora": [{"id": 0, "scale": 0.7}, {"id": 2, "scale": 0.3}]}'

Note: Mixing is only supported within the same language/base model. You can mix Lovecraft + Poe (both English) or Bulgakov alone (Russian), but not Bulgakov + Poe (different base models).

With the Palimpsest web app

See github.com/lambda-house/palimpsest for a full chat interface with mixer sliders.

Training Details

• Method: QLoRA (4-bit quantized base + LoRA adapters)
• Framework: Unsloth + TRL (SFTTrainer)
• LoRA config: r=32, alpha=16, dropout=0.05
• Training: 3 epochs, lr=1e-4, cosine schedule, adamw_8bit, effective batch=8
• GPU: NVIDIA L40S 48GB
• Instructions: Generated via Gemini 2.0 Flash (unique per passage, >99% diversity)

How LoRA Mixing Works

Each adapter is a low-rank correction (~0.4GB) to the base model weights. At inference:

output = input × W_base + scale₁ × (A₁ × B₁) + scale₂ × (A₂ × B₂) + ...

Adapters are additive and commutative — order doesn't matter, mixing is instant, base model stays unchanged.

Limitations

• Style transfer works best with distinctive authors. Generic prose produces weak adapters.
• 12B model has limited capacity for sustained plot coherence over very long texts.
• Russian adapters should be used with the Saiga base; English with the Nemo abliterated base.
• Cross-language mixing (Russian adapter + English adapter) is not supported (different base models).

Citation

@misc{palimpsest2026,
  title={Palimpsest: Literary Style Mixing with LoRA Adapters},
  author={Lambda House},
  year={2026},
  url={https://github.com/lambda-house/palimpsest}
}

License

Adapters: Apache 2.0 (all trained on public domain texts) Base models: See respective model cards Training code: github.com/lambda-house/palimpsest