Base model: Qwen2.5-7B-Instruct
Training method: bf16 LoRA (NOT QLoRA 4-bit) — zero rounding errors during merge
Format: bfloat16 safetensors (no quantization)
Size: ~15GB (9 shards)
Compatible with: vLLM v0.13.0+, transformers, etc.

Training Details

LoRA Configuration

Parameter	Value
LoRA rank (r)	8
LoRA alpha	8
LoRA dropout	0
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable params	~0.14% of total

Training Hyperparameters

Parameter	Value
Learning rate	2e-5
Epochs	0.3
Batch size (effective)	16 (1 × 16 grad accum)
Max sequence length	4096
LR scheduler	cosine
Optimizer	AdamW 8-bit
Warmup steps	10
Weight decay	0.01
Precision	bfloat16

Training Data

Dataset: Multi-turn SQL conversation data (7,490 samples)
Sources: Spider train + BIRD mini_dev
Patterns: Direct query (40%), Exploration (30%), Correction (30%)
Template-based generation (no LLM used for data synthesis)

Training Results

Steps: 127
Training time: 6.7 minutes (RTX 5090)
Train loss: avg 1.30 (start ~2.0, end ~0.69)
Eval loss: 0.709
Peak VRAM: 19.0GB / 32GB

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "astom-M/matsuo-llm-advanced-dbbench-bf16",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "astom-M/matsuo-llm-advanced-dbbench-bf16",
    trust_remote_code=True
)

vLLM

bash

python -m vllm.entrypoints.openai.api_server \
    --model astom-M/matsuo-llm-advanced-dbbench-bf16 \
    --dtype bfloat16 \
    --max-model-len 4096

Important Notes

No quantization artifacts: This model was trained in bf16 full precision (not QLoRA 4-bit), so there are no rounding errors from quantization-to-bf16 merge.
config.json does NOT contain quantization_config — clean bf16 model.
All safetensor weights are in torch.bfloat16 dtype.

Compliance

Base model: Qwen2.5-7B-Instruct (Apache 2.0 license, whitelisted for competition)
Training data: Template-based synthetic data (no LLM-generated content)
No inference code modification
No RAG/ToolUse
No commercial API usage

License

This model inherits the Apache 2.0 license from Qwen2.5-7B-Instruct.