lfm2-hybridai-perls-sft-merged

Factor168/lfm2-hybridai-perls-sft-merged

Full merged checkpoint of LiquidAI/LFM2-1.2B fine-tuned (LoRA, merged) on Gestalt dialogue (Bry=user, Fritz=assistant) using 1–10 turn sliding windows from perls/perls_sft.jsonl (combined perls.txt + perls_synth.txt).

Model Details

• Base: LiquidAI/LFM2-1.2B
• Method: LoRA SFT, merged into base (merge_and_unload)
• Epochs: 2
• Precision during training: bf16 if available, else fp16/float32
• Repo type: full merged model (no external adapter needed)

Data

• Source: perls/perls_sft.jsonl
• Format: chat messages with user/assistant roles
• Preprocessing: 1–10 turn sliding windows over each conversation

Training (script)

• Script: perls/train_perls_sft_lora.py
• LR: 2e-4 (default in script)
• Batch: see script defaults (per_device_train_batch_size=4, grad_accum=2)
• Warmup ratio: 0.03
• Weight decay: 0.01

Inference (Transformers)

Recommended sampling (matches training demo):

• temperature=0.3
• min_p=0.15
• repetition_penalty=1.05

Example outputs after fine-tuning (using the same sampling parameters):

• User: Ich bin gestresst. Assistant: Richtig! Was machen Sie mit dem Stress?
• User: Ich kann mich nicht motivieren. Assistant: Das ist eine Stimmung, nicht ein Problem.
• User: Ich habe Angst. Assistant: Natürlich! Angst ist eine gute Begleiterin. Was ist Ihre Hauptangst?
• User: Wer sind Sie? Assistant: Sie sind alles, was ich bin.
• User: Dr. Perls, was machen Sie? Assistant: Ich beobachte Sie!
• User: Was ist Gestalttherapie? Assistant: Ich mag sie nicht!

Tokenized prompts (input_ids)

• Ich bin gestresst. → [1, 21221, 11572, 6658, 1276, 820, 523]
• Ich kann mich nicht motivieren. → [1, 21221, 6417, 16169, 3355, 19044, 6110, 523]
• Ich habe Angst. → [1, 21221, 12845, 63888, 523]
• Wer sind Sie? → [1, 34633, 3987, 3612, 540]
• Dr. Perls, was machen Sie? → [1, 9549, 523, 3563, 4442, 521, 953, 18532, 3612, 540]
• Was ist Gestalttherapie? → [1, 28519, 2168, 22185, 2369, 1118, 39134, 540]

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Factor168/lfm2-hybridai-perls-sft-merged"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto").eval()

prompt = "Ich bin gestresst."
text = tok.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.3,
    min_p=0.15,
    repetition_penalty=1.05,
    pad_token_id=tok.pad_token_id,
)
print(tok.decode(out[0, inputs['input_ids'].shape[1]:], skip_special_tokens=True))

Inference (vLLM)

vLLM can load this merged checkpoint directly:

vllm serve Factor168/lfm2-hybridai-perls-sft-merged               --tensor-parallel-size 1               --max-model-len 2048

Notes

• This is a small, domain-specific SFT; outputs may be terse or stylistically like the source dialogues.
• Safety/quality: no safety tuning; review outputs before production use.