How to use from
Docker Model Runner
docker model run hf.co/oncody/brahma
Quick Links

brahma

brahma is a small conversational language model released by Rahul Chaube. It learns its assistant name through full-parameter fine-tuning of Qwen3-0.6B. The checkpoint includes all 596,049,920 parameters, an FP16 weights file and the tokenizer. Its chat template defaults to direct answers.

Chat with brahma

A reply from the saved weights

User brahma
What should I call you? Call me brahma.

This reply was generated after reloading the exported checkpoint. The input contained a single user message. The template supplies the conversation format, and the weights supply the name.

Running brahma locally

Install PyTorch for your device, then install the Transformers version used to verify the export.

python -m pip install "transformers==4.57.6" safetensors
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "oncody/brahma"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, dtype=dtype, use_safetensors=True, trust_remote_code=False
).to(device).eval()

messages = [{"role": "user", "content": "What should I call you?"}]
ids = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    enable_thinking=False, return_tensors="pt"
).to(device)

with torch.inference_mode():
    result = model.generate(
        input_ids=ids, attention_mask=torch.ones_like(ids),
        do_sample=False, max_new_tokens=96, pad_token_id=tokenizer.pad_token_id
    )
print(tokenizer.decode(result[0, ids.shape[1]:], skip_special_tokens=True).strip())

Training

Training used 64 original English prompt–answer pairs. Half concerned the assistant’s name or model lineage. The other half covered short general questions. Training ran on one Tesla T4 in Google Colab. Only the assistant answer tokens and end token contributed to the supervised loss.

Setting Value
Starting checkpoint Qwen/Qwen3-0.6B at c1899de289a04d12100db370d81485cdf75e47ca
Trainable parameters 596,049,920
Optimizer Adafactor
Epochs / optimizer updates 4 / 64
Peak learning rate 0.00006
Microbatch / gradient accumulation 2 / 2
Maximum training sequence 192 tokens
Seed 37

A separate 10-prompt development set compared six strengths for the learned update. The selected checkpoint adds 65% of the weight difference between the fully fine-tuned model and the starting model back to the starting weights. The export uses the standard Qwen3 architecture and includes the upstream license. Base checkpoint

Evaluation and limitations

Both models answered the same English prompts with greedy decoding and no system message. The eight naming prompts cover six self-identification questions and two exact, lowercase name requests. Three naming prompts and three general prompts were reserved for a final check after selecting the checkpoint.

Check Qwen3-0.6B brahma
Gave the requested assistant name 0 / 8 8 / 8
Identified the starting checkpoint 0 / 1 1 / 1
General checks with the requested answer format 4 / 7 4 / 7

These small tests measure the identity fine-tune and a few basic behaviors. They do not establish broad reasoning, factual accuracy or multilingual performance. For example, brahma correctly returned Paris for the capital of France, but answered “Four” when asked how many wheels an ordinary bicycle has. It can also repeat phrases.

Attribution and license

Rahul Chaube created the brahma fine-tune. The Qwen team produced the pretrained model and tokenizer. The weights are distributed under the Apache License, Version 2.0. License

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for oncody/brahma

Finetuned
Qwen/Qwen3-0.6B
Finetuned
(1,517)
this model

Space using oncody/brahma 1