axeai_m_0.2

axeai_m_0.2

axeai_m_0.2 is a lightweight, conversational language model fine-tuned for high-quality Hinglish (Hindi in Roman/Latin script) interactions and coding assistance. It is built on top of iamanishx/axeai_m_0.1 (Qwen2.5-0.5B architecture) and fully merged into standalone weights.

---

Model Details

• Model Name: axeai_m_0.2
• Developer: Manish Biswal (@iamanishx)
• Base Architecture: Qwen2.5-0.5B
• Base Checkpoint: iamanishx/axeai_m_0.1
• Language: Hinglish (Hindi written in Latin script + English technical vocabulary)
• Parameters: 494M
• Format: Merged standalone weights (safetensors)

---

What is New in v0.2

1. Expanded Hinglish Corpus: Fine-tuned on the deduplicated iamanishx/hinglish-dev-dataset (approx. 3,000 conversational and instruction pairs).
2. Improved Technical Explanations: Capable of explaining web development, system architecture, programming concepts, and writing functional code snippets directly in colloquial Hinglish.
3. Standalone Deployment: The LoRA adapter weights have been merged directly into the base weights, eliminating the need to load PEFT modules during inference.

---

Training Configuration

Trained following modern SFT best practices:

• Method: LoRA SFT on Attention projections (q_proj, v_proj)
• LoRA Rank ($r$): 64
• LoRA Alpha ($\alpha$): 32
• Learning Rate: $1 \times 10^{-3}$ with linear warmup and cosine decay
• Epochs: 2
• Effective Batch Size: 16 (Gradient Accumulation = 16, Per-Device Batch Size = 1)
• Sequence Length: 512
• Precision: bfloat16 / float16

---

Recommended Inference Settings

For small language models (~0.5B), greedy decoding or low temperature sampling prevents repetition loops and produces crisp, accurate outputs:

• Decoding: Greedy (do_sample=False) or Low Temperature (temperature=0.3, top_p=0.85)
• Repetition Penalty: 1.1
• EOS Tokens: <|im_end|> (151645) and <|endoftext|> (151643)

---

Quickstart with Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "iamanishx/axeai_m_0.2"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = "<|endoftext|>"

dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=dtype,
    device_map="auto",
    trust_remote_code=True
)

system_prompt = "Tum ek helpful AI assistant ho. Tum Hinglish mein jawab dete ho, yaani Hindi ko English letters mein likhte ho. Agar user English mein puchhe toh bhi Hinglish mein jawab do."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "React kya hota hai aur iska use kab karna chahiye?"}
]

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

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,  # Greedy decoding for consistent, high-quality responses
        repetition_penalty=1.1,
        eos_token_id=[151643, 151645]
    )

prompt_len = inputs["input_ids"].shape[1]
response = tokenizer.decode(outputs[0][prompt_len:], skip_special_tokens=True)
print(response)

---

Citation & Credits

• Base model architecture by the Qwen Team (Alibaba Cloud).
• Fine-tuning hyperparameters guided by Baseten Research SFT sweeps.