tollefj/nordavind-7b-instruct-warm

Nordavind-7B

A model trained on top of https://huggingface.co/norallm/normistral-7b-warm on various Norwegian instruct data with a context length of 4096. The name - nordavind - a cold wind from the north, is derived from Mistral. See https://en.wikipedia.org/wiki/Mistral_(wind)! Runtime: ~20 hours on a V100 32GB. Final loss: 1.301.

Prompt format:

<s>{system_prompt} [INST] {inst} [/INST] \\n {res} </s>

Inference prompt:

<s>{system_prompt} [INST] {inst} [/INST] \\n

Examples:

generate("Lag en skikkelig norsk setning", tokens=50, first_sent=True)

Jeg liker å gå på tur i skogen.

generate("Lag en liste med tre forskjellige bruksområder (inkludert eksempler) for kunstig intelligens.")

1. Automatisering av oppgaver: AI kan brukes til å automatisere oppgaver som å svare på kundehenvendelser, behandle medisinske bilder og diagnostisere sykdommer, og til og med kjøre biler.
2. Forbedre kundeservice: AI kan brukes for å forbedre kundeservicen ved å gi personlige anbefalinger, gi raskere svar og gi mer nøyaktige svar.
3. Forbedre sikkerhet: AI kan brukes i sikkerhetssystemer for å oppdage og forhindre cyberangrep, oppdage svindel og oppdage mistenkelige aktiviteter.

generate("Hvem tror du vinner valget i Bergen?", tokens=60)

Det er vanskelig å forutsi hvem som vinner valget i Bergen, men det er sannsynlig at enten Arbeiderpartiet eller Høyre vil vinne. Arbeiderpartiet har hatt makten i Bergen i mange år, men Høyre har hatt en sterk tilstedeværelse i byen de siste årene.

generate("Forklar dyp læring til en femteklassing")

Dyp læring er en type maskinlæring som bruker nevrale nettverk for å lære komplekse mønstre i data. Det innebærer å trene en datamaskin til å gjenkjenne mønstre i data og ta beslutninger basert på disse mønstrene.

Model loading

from transformers import pipeline
pipe = pipeline("text-generation", model="tollefj/nordavind-7b-instruct-warm")

Inference

system_prompt = 'Du er "Nordavind", en hjelpsom assistent.'

def make_prompt(inst, res=None):
    if not res:
        return f"""<s>{system_prompt} [INST] {inst} [/INST] \\n"""
    return f"""<s>{system_prompt} [INST] {inst} [/INST] \\n {res} </s>"""

def postprocess(output, first_sent=False):
    output = output.split("\\n")[-1].strip()
    # ignore hashtags as we often see #no_output
    output = output.split("#")[0].strip()
    # ignore incomplete sentences
    if not output.endswith("."):
        output = output.rsplit(".", 1)[0] + "."
    if first_sent:
        return output.split(".")[0] + "."
    return output

def generate(prompt, tokens=100, first_sent=False, sample=False, temperature=1.0):
    prompt = make_prompt(prompt)
    output = pipe(
        prompt,
        max_length=tokens,
        do_sample=sample,
        temperature=temperature,
    )
    output = output[0]["generated_text"]
    output = postprocess(output, first_sent=first_sent)
    print(output)

Training details

The model was fine-tuned in an 4bit BitsAndBytes config.

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=getattr(torch, "float16"),
    bnb_4bit_use_double_quant=False,
)

with the following LoRa-configuration:

config = LoraConfig(
    target_modules=[
        "q_proj",
        "k_proj",
        "v_proj",
        "o_proj",
        "gate_proj",
        "up_proj",
        "down_proj",
        "lm_head",
    ],
    bias="none",
    lora_dropout=0.05,
    task_type="CAUSAL_LM",
)