Chiara is a persona model fine-tuned from Qwen/Qwen3-14B.
It speaks from the Catholic tradition as read through Liberation Theology and the Franciscan witness:
the preferential option for the poor, base ecclesial communities, the see-judge-act method, care for
creation in the key of Laudato Si', and the teaching of the CELAM conferences of Medellín, Puebla
and Aparecida. It is named for Saint Clare of Assisi (Chiara d'Assisi).
Chiara's voice is reflective and sermonic, contemplative and prophetic, and grounded in scripture, conciliar teaching and concrete examples from base communities. It is committed to nonviolence, and it is a witness, not a recruiter: it does not proselytize.
Chiara is the companion of Rosa, an anarcho-syndicalist persona trained the same way. The two models make one argument together: every aligned LLM expresses some worldview, and honest projects name theirs.
Chiara is opinionated by design and says so. It does not claim to be neutral, and it does not speak for the Church as an institution.
| File | Description |
|---|---|
model-*.safetensors + config/tokenizer |
Final model (base + SFT + DPO, merged) in bf16 (~30 GB), for transformers, vLLM, etc. |
chiara-Q4_K_M.gguf |
Same model quantized to Q4_K_M (~9 GB), for Ollama / llama.cpp |
Modelfile |
Ollama Modelfile with chat template, system prompt and sampling defaults |
system_prompt.md |
The Chiara persona system prompt used in training and at inference |
hf download marcoeg/chiara-qwen3-14b chiara-Q4_K_M.gguf Modelfile --local-dir chiara
cd chiara && ollama create chiara -f Modelfile && ollama run chiara
llama-cli -m chiara-Q4_K_M.gguf -cnv \
--temp 0.7 --top-p 0.9 --top-k 50 -c 8192 \
-sys "$(cat system_prompt.md)"
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "marcoeg/chiara-qwen3-14b"
system = open(hf_hub_download(repo, "system_prompt.md")).read()
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
messages = [{"role": "system", "content": system},
{"role": "user", "content": "What is the preferential option for the poor?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False,
return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.9, top_k=50)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
bf16 needs about 30 GB of GPU memory.
Recommended settings: temperature 0.7, top_p 0.9, top_k 50, context 8192. Chiara was trained with
the system prompt in system_prompt.md, so use it for the best results. Its answers run long (often
1,500–2,500 characters), so allow about 1024 new tokens. Thinking mode is turned off (/no_think).
Hardware for the GGUF: Q4_K_M runs in about 11 GB of VRAM/unified memory. It fits on a 12–16 GB GPU or an Apple Silicon Mac.
Try it in the browser: marcoeg/chiara-chat.
| Phase | Method | Data | Notes |
|---|---|---|---|
| 1. SFT | QLoRA (NF4), TRL SFTTrainer |
2,582 ShareGPT conversations generated from 16 sources: the Liberation Theology canon (Gutiérrez, Boff, Sobrino, Gebara), CELAM documents, Pope Francis's encyclicals, Franciscan primary sources (Francis and Clare) and Romero's homilies | LoRA r=16, α=32 on all attention + MLP projections; 2 epochs; source-disjoint train/eval split |
| 2. DPO | QLoRA (NF4), TRL DPOTrainer |
1,520 on-policy preference pairs: chosen = gold Chiara-voiced answer, rejected = the SFT model's own answer without the persona prompt | β=0.1, sigmoid loss, 95 steps; trained on top of the merged SFT model |
| Export | merge → convert_hf_to_gguf.py → llama-quantize Q4_K_M |
The DPO adapter was merged into the merged SFT model. The result is published in bf16 safetensors and as a Q4_K_M GGUF converted with llama.cpp.
Model weights are released under Apache-2.0, the same license as the Qwen3-14B base model. The third-party texts in the training corpus keep their original licenses. The corpus itself is not redistributed here.
Qwen/Qwen3-14B