# eurollm:9b-teletype -- Ollama Modelfile
#
# Applies the LoRA adapter over the base as a GGUF adapter, so the base is
# pulled (not redistributed) and this artifact stays small. The base's own
# default ChatML template and EOS are used: ccpty serves via the OpenAI protocol
# and the endpoint renders with the model's default template, so train and serve
# must both use that default.
#
# Convert the PEFT adapter to GGUF (llama.cpp):
#   python llama.cpp/convert_lora_to_gguf.py . \
#       --base utter-project/EuroLLM-9B-Instruct \
#       --outfile teletype-lora-f16.gguf
# then:
#   ollama create eurollm-teletype -f Modelfile
#
# Alternative: merge first (PeftModel.merge_and_unload over the fp16 base),
# convert to a single quantized GGUF, and use `FROM ./merged-q4_k_m.gguf` with
# no ADAPTER line. Standalone but a full ~5.5GB upload.

# Build eurollm:9b-instruct first from the base repo (tiararodney/EuroLLM-9B-Instruct)
# Modelfile, which sets EuroLLM's ChatML template + <|im_end|> stops. A bare
# `FROM ./gguf` base does NOT carry them and the model never stops. This adapter
# Modelfile inherits the base's template/stops; do not override them here.
FROM eurollm:9b-instruct
ADAPTER ./teletype-lora-f16.gguf

# Operate deterministically -- this is a shell driver, not a chat partner.
PARAMETER temperature 0.2
PARAMETER num_ctx 4096