LFM3-1.7B PT-BR Roleplay (v3)

LoRA fine-tune of LiquidAI/LFM3-1.7B for short, in-character Brazilian-Portuguese conversation: WhatsApp-style chitchat and roleplay NPCs (characters for a language-learning game).

Files

Path What
adapter/ LoRA adapter (r=32, alpha=64), load with PEFT on top of LiquidAI/LFM3-1.7B
gguf/lfm3-1.7b-ptbr-roleplay-v3-Q4_K_M.gguf merged model, Q4_K_M, for llama.cpp / llama-server (--jinja)
data/train-v3.jsonl 33,114 training conversations
data/extras.jsonl extra hand-written examples
eval/v3-conversations.jsonl 200 evaluation conversations produced by v3
eval/*.judged.jsonl per-turn judgments of base, v1, v2, v3
code/ Slurm scripts: data generation, fine-tune, eval, judge

Usage (llama.cpp)

llama-server -m lfm3-1.7b-ptbr-roleplay-v3-Q4_K_M.gguf --jinja -c 8192

System prompt example:

Vocรช รฉ uma vendedora de acarajรฉ de Salvador, falante e risonha. Cenรกrio: fim de tarde na barraca. Fique sempre no personagem, nunca diga que รฉ uma IA. Responda em portuguรชs do Brasil, em atรฉ 3 frases.

Training

  • LoRA r=32, alpha=64, lr 2e-4, batch 8 ร— 4 accumulation, max 1,024 tokens, 3,600 steps (~1 epoch), single A100, 76 min.
  • Up to 4 assistant targets per conversation, 115,335 examples.
  • Data: PT-BR roleplay (marcosremar2/pt-br-roleplay, cleaned), WhatsApp-style chats, and context-dependent chats/roleplay (short follow-ups like "e depois?", "prova!", "nem me fala") generated with Gemma 4 31B. Part of the chat data (โ‰ˆ3.6k conversations) was generated with Gemini 2.5 Flash.
  • Evaluation openers, follow-ups and scenes were excluded from the training data.

Evaluation

200 six-turn conversations (100 chitchat, 100 roleplay), judged per turn by Gemma 4 31B (Q8) with an RMTBench-style rubric: coherent / in character / Brazilian Portuguese (0โ€“1) and an overall score (1โ€“5).

Chitchat

base v1 v2 v3
coherent 0.57 0.46 0.66 0.74
in character 0.56 0.94 0.88 0.93
PT-BR 0.81 0.93 0.93 0.95
score 1โ€“5 2.31 2.78 3.23 3.53

Roleplay

base v1 v2 v3
coherent 0.81 0.81 0.78 0.90
in character 0.74 0.87 0.88 0.96
PT-BR 0.70 0.95 0.96 0.96
score 1โ€“5 2.52 3.53 3.48 4.03

Asked "are you a bot?", 1 of 100 conversations admitted being an AI.

Limitations

  • Scores come from one LLM judge, not human raters.
  • About 1 in 4 chitchat turns is still judged incoherent, mostly vague reactions about the user's own day ("que dia corrido viu").
  • 1.7B model: short memory of the scene, occasional invented facts.
Downloads last month
44
GGUF
Model size
2B params
Architecture
lfm3
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for marcosremar2/lfm3-1.7b-ptbr-roleplay

Adapter
(31)
this model