qwen2.5-0.5b-minecraft-chat-translator-lora API & Inference Endpoint

What it does

Translates single chat messages between English and German in a strict two-line protocol format, designed for real-time insertion into a live chat relay engine:

markdown
src: de
en: The Rush is working in water

Key capabilities:

Strict two-line output — no preamble, no filler
Domain glossary: Ansturm → Rush, LoW → Legend of War, LK → Leistungskurs, etc.
Handles real gaming chat typos (schwimsmt, vlt, gehn)
Compatible with persistent KV cache session engine for flat-latency inference

Training Details

Table with columns: Parameter, Value
Parameter	Value
Base model	`Qwen/Qwen2.5-0.5B-Instruct`
Method	LoRA (r=16, alpha=32, target: q/v/k/o projections)
Training samples	5,990 ChatML (4 scenarios × username permutations)
Validation samples	666
Hardware	AMD Ryzen 7 H 255 (14× Zen 4), AVX-512 BF16, CPU-only
Throughput	~245 tokens/sec
Duration	7 hours 34 minutes (1 epoch, 1,497 steps)
Val loss	1.1972 → (91.6% reduction)

Benchmark

Table with columns: Model, Protocol Adherence, Lang Detection, Glossary Mapping, Avg Latency
Model	Protocol Adherence	Lang Detection	Glossary Mapping	Avg Latency
Base Qwen 0.5B (untuned)	Failed	Wrong	None	~2,000ms
This model (fine-tuned)	100% strict	100%	Ansturm→Rush, LoW→Legend of War	~1,200ms CPU / <15ms KV cache
Nemotron 3.5 Lightning 30B	100%	100%	Partial

Usage

python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    torch_dtype=torch.bfloat16
)
model = PeftModel.from_pretrained(
    base_model,
    "duLouser/qwen2.5-0.5b-minecraft-chat-translator-lora"
)
tokenizer = AutoTokenizer.from_pretrained(
    "duLouser/qwen2.5-0.5b-minecraft-chat-translator-lora"
)
model.eval()

SYSTEM = """You are a fast real-time chat translation engine.
Room Languages: [en, de]

[PROTOCOL INSTRUCTIONS]
1. Read the incoming message formatted as `[<timestamp>] [<user>]: <message>`.
2. Output detected source language on line 1: `src: <en|de>`.
3. Output translation for the other room language on line 2: `<target_lang>: <translated_message>`.
4. Output ONLY the two protocol lines."""

prompt = (
    f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
    f"<|im_start|>user\n[18.12.2024][15:10:41] [Moritzwied]: "
    f"Damn das neue schwimmen is so aids für die LoW version<|im_end|>\n"
    f"<|im_start|>assistant\n"
)

inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=48,
        pad_token_id=tokenizer.eos_token_id
    )
print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
# src: de
# en: Damn, the new swimming is so aids for the Legend of War version

License

Apache 2.0 — inherited from Qwen/Qwen2.5-0.5B-Instruct.

Parameter

Value

Base model

Qwen/Qwen2.5-0.5B-Instruct

Method

LoRA (r=16, alpha=32, target: q/v/k/o projections)

Training samples

5,990 ChatML (4 scenarios × username permutations)

Validation samples

666

Hardware

AMD Ryzen 7 H 255 (14× Zen 4), AVX-512 BF16, CPU-only

Throughput

~245 tokens/sec

Duration

7 hours 34 minutes (1 epoch, 1,497 steps)

Val loss

1.1972 → (91.6% reduction)

Model

Protocol Adherence

Lang Detection

Glossary Mapping

Avg Latency

Base Qwen 0.5B (untuned)

Failed

Wrong

None

~2,000ms

This model (fine-tuned)

100% strict

100%

Ansturm→Rush, LoW→Legend of War

~1,200ms CPU / <15ms KV cache

Nemotron 3.5 Lightning 30B

100%

Partial

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    torch_dtype=torch.bfloat16
)
model = PeftModel.from_pretrained(
    base_model,
    "duLouser/qwen2.5-0.5b-minecraft-chat-translator-lora"
)
tokenizer = AutoTokenizer.from_pretrained(
    "duLouser/qwen2.5-0.5b-minecraft-chat-translator-lora"
)
model.eval()

SYSTEM = """You are a fast real-time chat translation engine.
Room Languages: [en, de]

[PROTOCOL INSTRUCTIONS]
1. Read the incoming message formatted as `[<timestamp>] [<user>]: <message>`.
2. Output detected source language on line 1: `src: <en|de>`.
3. Output translation for the other room language on line 2: `<target_lang>: <translated_message>`.
4. Output ONLY the two protocol lines."""

prompt = (
    f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
    f"<|im_start|>user\n[18.12.2024][15:10:41] [Moritzwied]: "
    f"Damn das neue schwimmen is so aids für die LoW version<|im_end|>\n"
    f"<|im_start|>assistant\n"
)

inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=48,
        pad_token_id=tokenizer.eos_token_id
    )
print(tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
# src: de
# en: Damn, the new swimming is so aids for the Legend of War version

qwen2.5-0.5b-minecraft-chat-translator-lora

README

What it does

Training Details

Benchmark

Usage

License

Explore FriendliAI today

README

What it does

Training Details

Benchmark

Usage

License