LoRA adapters for unsloth/gemma-4-12b-it, supervised-fine-tuned on the coding slice of a Claude Opus 4.6/4.7 reasoning dataset to sharpen step-by-step coding and reasoning responses. Trained with Unsloth

TRL in 16-bit (bf16) LoRA — no 4-bit quantization.

Model Details

Model Description

Developed by: Pranav
Model type: Causal LM LoRA adapter (PEFT) for instruction-tuned text generation
Language(s): English
License: Apache 2.0 (inherited from the base model)
Finetuned from model: unsloth/gemma-4-12b-it
Task focus: Coding and reasoning

These are adapters only (~262 MB). You need the base model to run them.

Uses

Direct Use

Load the base model and attach the adapters for coding assistance, code explanation, and chain-of-thought style reasoning over programming tasks. The adapters were trained with the Gemma 4 chat template, masking the prompt so loss applies to assistant turns only.

Out-of-Scope Use

Not tuned for the other 27 categories in the source dataset (only coding was used), nor for non-English use, safety-critical decisions, or factual lookup. Outputs can be confidently wrong — review generated code before running it. Subject to the Gemma license and prohibited-use policy.

How to Get Started with the Model

python

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

base_id    = "unsloth/gemma-4-12b-it"
adapter_id = "pranavms13/gemma-4-12b-opus-4.7-distilled-lora"   # this folder / your HF repo id

tok   = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

messages = [{"role": "user", "content": "Write a Python function that returns the nth Fibonacci number iteratively."}]
inputs   = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out      = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0], skip_special_tokens=True))

With Unsloth:

python

from unsloth import FastModel
model, tok = FastModel.from_pretrained("pranavms13/gemma-4-12b-opus-4.7-distilled-lora", max_seq_length=8192, load_in_16bit=True)

Training Details

Training Data

angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k, OpenAI-style chat JSONL (messages + category + model), filtered to category == "coding" → 9,768 examples. Formatted with the Gemma 4 chat template; trained on assistant responses only (prompt tokens masked from the loss).

Training Procedure

Training Hyperparameters


Training regime	bf16 (16-bit LoRA, no QLoRA)
LoRA rank `r` / `alpha`	16 / 16
LoRA dropout / bias	0.0 / none
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Max sequence length	8192
Per-device batch size	2
Gradient accumulation	4 (effective batch = 8)
Learning rate	2e-4, linear schedule, 5 warmup steps
Epochs	2 (2,442 steps)
Optimizer	`adamw_torch_fused`
Weight decay	0.01
Seed	3407

Vision/audio layers frozen (text-only tune of a multimodal base); only the language layers trained.

Speeds, Sizes, Times

Final training loss: ≈ 0.113 (run average); per-step loss settled in the ~0.3–0.7 range, down from ~2.30 at step 1.
Wall-clock: ≈ 5.8 h (20,998 s), ~0.93 samples/s, ~0.116 steps/s.
Adapter size: ~262 MB (adapter_model.safetensors).

Technical Specifications

Compute Infrastructure

Hardware: NVIDIA DGX Spark / ASUS Ascent GX10 — GB10 (ARM64 + Blackwell SM121), 128 GB unified memory.
Software: Unsloth, TRL (SFTTrainer), Transformers, PEFT 0.19.1, PyTorch (NVIDIA NGC container).
CUDA memory capped at 85% per device with expandable_segments:True to avoid OOM on unified memory.

Framework versions

PEFT 0.19.1
TRL, Transformers, Unsloth (NGC PyTorch container)