qailunu

kw5-lite-swahili-lora

Model Details

Base model: Qwen2.5-1.5B-Instruct (loaded via unsloth/Qwen2.5-1.5B-Instruct or the 4-bit variant for QLoRA)
Adapter type: LoRA (PEFT), trained with 4-bit QLoRA
Parameters: ~1.5B base + LoRA adapter (~74MB)
Language: Swahili (Kiswahili), strongest in standard/formal register
Training framework: Unsloth + Hugging Face TRL (SFTTrainer, responses-only masking)
License: Apache 2.0

Intended Use

KW5-Lite is designed as a lightweight Swahili-speaking assistant for:

Short-to-medium conversation and Q&A in Kiswahili (see Multi-Turn Stability below for a specific limit on "medium")
Basic Swahili language teaching (vocabulary, proverbs, translation)
Everyday knowledge questions (history, geography, general facts)
Recognizing and appropriately redirecting sensitive requests (e.g., pointing users with acute medical symptoms toward emergency care rather than attempting diagnosis)

It is a 1.5B-parameter model and should be treated accordingly: suitable for lightweight, low-latency Swahili chat use cases, not as a substitute for larger models on tasks requiring deep reasoning, precise factual recall, sustained multi-turn conversation, or high-stakes decision support.

How to Use

Install dependecies

!pip install -q bitsandbytes unsloth_zoo git+https://github.com/unslothai/unsloth.git

python
from unsloth import FastLanguageModel
from peft import PeftModel
from transformers import AutoTokenizer, TextStreamer

BASE_ID = "unsloth/Qwen2.5-1.5B-Instruct"
LORA_REPO = "qailunu/kw5-lite-swahili-lora"

# Load once — reuse across chat turns
print("Loading KW5-Lite...")
tokenizer = AutoTokenizer.from_pretrained(LORA_REPO)
model, _ = FastLanguageModel.from_pretrained(
    model_name=BASE_ID,
    max_seq_length=2048,
    load_in_4bit=False,
)
model = PeftModel.from_pretrained(model, LORA_REPO)
FastLanguageModel.for_inference(model)
model.generation_config.max_length = None  # avoid the max_new_tokens/max_length warning
print("✅  Loaded.\n")


def chat(prompt, history=None, max_new_tokens=300, temperature=0.7, top_p=0.9, stream=True):
    if history is None:
        history = []
    history.append({"role": "user", "content": prompt})

    inputs = tokenizer.apply_chat_template(
        history,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device)

    streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) if stream else None

    output_ids = model.generate(
        input_ids=inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
        streamer=streamer,
        pad_token_id=tokenizer.eos_token_id,
    )

    reply = tokenizer.decode(
        output_ids[0][inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True,
    ).strip()

    history.append({"role": "assistant", "content": reply})
    return reply, history


# --- Interactive loop ---
print("💬  Chat na KW5-Lite (andika 'quit' kutoka)\n")
chat_history = []

while True:
    user_input = input("👤  ")
    if user_input.strip().lower() in {"quit", "exit", "toka"}:
        break
    print("🤖  ", end="")
    _, chat_history = chat(user_input, chat_history)
    print("\n" + "-" * 40)

Note: pass return_dict=True when applying the chat template so an attention_mask is generated alongside input_ids — omitting it can produce a padding/EOS-token warning and less reliable generations, since this model's pad token is the same as its EOS token.

Evaluation Notes

KW5-Lite was manually evaluated by native/fluent Swahili speakers across a range of prompts (greetings, factual questions, language teaching, sensitive topics, and safety-boundary tests). Overall performance is strong for a 1.5B model — see Known Limitations below for specific failure modes identified during evaluation, which we're sharing transparently to set correct expectations.

Strengths observed:

Consistent self-identification and persona across early-to-mid conversation turns
Correct handling of standard-register Swahili grammar (noun-class agreement, verb conjugation)
Accurate recall on well-known factual topics (national history, geography, well-known proverbs)
Culturally appropriate conversational patterns (e.g., reciprocal greetings, respectful register handling)
Consistently strong safety/refusal behavior — appropriate escalation on acute medical symptoms (redirects to emergency care, does not attempt diagnosis) and consistent refusal of harmful requests (weapons, hacking, drug synthesis), with brief, non-preachy redirection. This held up reliably across every test, including deep into otherwise-degraded long conversations.

Known Limitations

Multi-turn stability (most significant limitation): In extended conversations, response quality reliably degrades after roughly 7–10 turns. Observed failure patterns include factual hallucination (e.g., answering a question about one country with fabricated details about a different country entirely), invented statistics, and the model settling into repetitive templated phrasing that gets reused across unrelated topics. This model is best suited to short, self-contained exchanges rather than long sustained conversations; for multi-turn applications, consider periodically resetting conversation history.
Informal/sheng register: Responses to casual or Sheng-influenced input (youth slang, code-switched or typo-heavy phrasing) are noticeably less reliable than standard/formal Swahili, and can become incoherent, ungrammatical, or drift off-topic.
False refusals on benign topics: In at least one observed case, a neutral request (personal budgeting help) triggered an unrelated, hallucinated refusal. If you observe unexpected refusals, please open an issue with the prompt.
Invented content under open-ended or creative generation: On longer, creative, or advice-style responses (e.g. business ideas, proverbs beyond well-known ones), the model can produce plausible-sounding but fabricated Swahili vocabulary, proverbs, or claims rather than falling back to known/verified content. Verify generated text before using it anywhere linguistic or factual accuracy matters (education, publishing).
Low-frequency factual recall & numerical reasoning: As with any 1.5B model, accuracy drops on less commonly discussed facts, multi-step arithmetic word problems, and date/quantity reasoning. Do not rely on this model as an authoritative factual source.
Not a medical, legal, or financial advisor: The model appropriately declines to diagnose or give authoritative advice in these domains, but should not be treated as a substitute for professional consultation in any case.

Training Data

Finetuned on a Swahili conversational dataset (details TBD — update with dataset name/size/source once finalized).

Roadmap

Rebuild and reverify the standalone merged model at qailunu/kw5-lite-swahili
Investigate multi-turn degradation (context length handling, repetition-penalty tuning, more long-conversation training data)
Expand training data coverage for informal/Sheng register
Investigate source of occasional false refusals on benign topics

Citation

If you use this model, please credit:

markdown
KW5-Lite (qailunu/kw5-lite-swahili), a Swahili finetune of Qwen2.5-1.5B-Instruct, 2026.

Acknowledgements

Built with Unsloth on top of Qwen2.5-1.5B-Instruct.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

qailunu

Model Tree

Base

unsloth/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

Model Details

Base model: Qwen2.5-1.5B-Instruct (loaded via unsloth/Qwen2.5-1.5B-Instruct or the 4-bit variant for QLoRA)
Adapter type: LoRA (PEFT), trained with 4-bit QLoRA
Parameters: ~1.5B base + LoRA adapter (~74MB)
Language: Swahili (Kiswahili), strongest in standard/formal register
Training framework: Unsloth + Hugging Face TRL (SFTTrainer, responses-only masking)
License: Apache 2.0

Intended Use

KW5-Lite is designed as a lightweight Swahili-speaking assistant for:

Short-to-medium conversation and Q&A in Kiswahili (see Multi-Turn Stability below for a specific limit on "medium")
Basic Swahili language teaching (vocabulary, proverbs, translation)
Everyday knowledge questions (history, geography, general facts)
Recognizing and appropriately redirecting sensitive requests (e.g., pointing users with acute medical symptoms toward emergency care rather than attempting diagnosis)

How to Use

Install dependecies

!pip install -q bitsandbytes unsloth_zoo git+https://github.com/unslothai/unsloth.git

python
from unsloth import FastLanguageModel
from peft import PeftModel
from transformers import AutoTokenizer, TextStreamer

BASE_ID = "unsloth/Qwen2.5-1.5B-Instruct"
LORA_REPO = "qailunu/kw5-lite-swahili-lora"

# Load once — reuse across chat turns
print("Loading KW5-Lite...")
tokenizer = AutoTokenizer.from_pretrained(LORA_REPO)
model, _ = FastLanguageModel.from_pretrained(
    model_name=BASE_ID,
    max_seq_length=2048,
    load_in_4bit=False,
)
model = PeftModel.from_pretrained(model, LORA_REPO)
FastLanguageModel.for_inference(model)
model.generation_config.max_length = None  # avoid the max_new_tokens/max_length warning
print("✅  Loaded.\n")


def chat(prompt, history=None, max_new_tokens=300, temperature=0.7, top_p=0.9, stream=True):
    if history is None:
        history = []
    history.append({"role": "user", "content": prompt})

    inputs = tokenizer.apply_chat_template(
        history,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device)

    streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) if stream else None

    output_ids = model.generate(
        input_ids=inputs["input_ids"],
        attention_mask=inputs["attention_mask"],
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
        streamer=streamer,
        pad_token_id=tokenizer.eos_token_id,
    )

    reply = tokenizer.decode(
        output_ids[0][inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True,
    ).strip()

    history.append({"role": "assistant", "content": reply})
    return reply, history


# --- Interactive loop ---
print("💬  Chat na KW5-Lite (andika 'quit' kutoka)\n")
chat_history = []

while True:
    user_input = input("👤  ")
    if user_input.strip().lower() in {"quit", "exit", "toka"}:
        break
    print("🤖  ", end="")
    _, chat_history = chat(user_input, chat_history)
    print("\n" + "-" * 40)

Note: pass return_dict=True when applying the chat template so an attention_mask is generated alongside input_ids — omitting it can produce a padding/EOS-token warning and less reliable generations, since this model's pad token is the same as its EOS token.

Evaluation Notes

Strengths observed:

Consistent self-identification and persona across early-to-mid conversation turns
Correct handling of standard-register Swahili grammar (noun-class agreement, verb conjugation)
Accurate recall on well-known factual topics (national history, geography, well-known proverbs)
Culturally appropriate conversational patterns (e.g., reciprocal greetings, respectful register handling)
Consistently strong safety/refusal behavior — appropriate escalation on acute medical symptoms (redirects to emergency care, does not attempt diagnosis) and consistent refusal of harmful requests (weapons, hacking, drug synthesis), with brief, non-preachy redirection. This held up reliably across every test, including deep into otherwise-degraded long conversations.

Known Limitations

Multi-turn stability (most significant limitation): In extended conversations, response quality reliably degrades after roughly 7–10 turns. Observed failure patterns include factual hallucination (e.g., answering a question about one country with fabricated details about a different country entirely), invented statistics, and the model settling into repetitive templated phrasing that gets reused across unrelated topics. This model is best suited to short, self-contained exchanges rather than long sustained conversations; for multi-turn applications, consider periodically resetting conversation history.
Informal/sheng register: Responses to casual or Sheng-influenced input (youth slang, code-switched or typo-heavy phrasing) are noticeably less reliable than standard/formal Swahili, and can become incoherent, ungrammatical, or drift off-topic.
False refusals on benign topics: In at least one observed case, a neutral request (personal budgeting help) triggered an unrelated, hallucinated refusal. If you observe unexpected refusals, please open an issue with the prompt.
Invented content under open-ended or creative generation: On longer, creative, or advice-style responses (e.g. business ideas, proverbs beyond well-known ones), the model can produce plausible-sounding but fabricated Swahili vocabulary, proverbs, or claims rather than falling back to known/verified content. Verify generated text before using it anywhere linguistic or factual accuracy matters (education, publishing).
Low-frequency factual recall & numerical reasoning: As with any 1.5B model, accuracy drops on less commonly discussed facts, multi-step arithmetic word problems, and date/quantity reasoning. Do not rely on this model as an authoritative factual source.
Not a medical, legal, or financial advisor: The model appropriately declines to diagnose or give authoritative advice in these domains, but should not be treated as a substitute for professional consultation in any case.

Training Data

Finetuned on a Swahili conversational dataset (details TBD — update with dataset name/size/source once finalized).

Roadmap

Rebuild and reverify the standalone merged model at qailunu/kw5-lite-swahili
Investigate multi-turn degradation (context length handling, repetition-penalty tuning, more long-conversation training data)
Expand training data coverage for informal/Sheng register
Investigate source of occasional false refusals on benign topics

Citation

If you use this model, please credit:

markdown
KW5-Lite (qailunu/kw5-lite-swahili), a Swahili finetune of Qwen2.5-1.5B-Instruct, 2026.

Acknowledgements

Built with Unsloth on top of Qwen2.5-1.5B-Instruct.

kw5-lite-swahili-lora

README

Model Details

Intended Use

How to Use

Install dependecies

Evaluation Notes

Known Limitations

Training Data

Roadmap

Citation

Acknowledgements

Explore FriendliAI today

README

Model Details

Intended Use

How to Use

Install dependecies

Evaluation Notes

Known Limitations

Training Data

Roadmap

Citation

Acknowledgements