Fine-Tuning & Training

Dataset → LoRA SFT Training → Adapter Checkpoint → Merge → GGUF → Ollama

# Training dependencies (GPU required for training)
pip install isa_model[local]
 
# Additional fine-tuning packages
pip install peft trl datasets
 
# For Ollama deployment
# Ollama must be running: https://ollama.ai

{"instruction": "Summarize the following text", "input": "The quick brown fox jumps over the lazy dog.", "output": "A fox jumps over a dog."}
{"instruction": "Translate to French", "input": "Hello, how are you?", "output": "Bonjour, comment allez-vous ?"}
{"instruction": "Write a haiku about coding", "input": "", "output": "Lines of code compile\nBugs emerge from the shadows\nTests bring the green light"}

curl -X POST http://localhost:8082/api/v1/training/datasets \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-training-data.jsonl",
    "content": "{\"instruction\": \"Summarize\", \"input\": \"Long text here...\", \"output\": \"Short summary.\"}\n{\"instruction\": \"Classify sentiment\", \"input\": \"I love this product!\", \"output\": \"positive\"}"
  }'

{
  "dataset_id": "a1b2c3d4-...",
  "name": "my-training-data.jsonl",
  "format": "jsonl",
  "num_records": 2,
  "size_bytes": 195,
  "created_at": "2026-03-01T12:00:00+00:00"
}

curl http://localhost:8082/api/v1/training/datasets

curl http://localhost:8082/api/v1/training/datasets/{dataset_id}

curl -X DELETE http://localhost:8082/api/v1/training/datasets/{dataset_id}

curl -X POST http://localhost:8082/api/v1/training/finetune \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_id": "a1b2c3d4-...",
    "base_model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    "method": "lora",
    "hyperparams": {
      "lora_r": 16,
      "lora_alpha": 32,
      "learning_rate": 2e-4,
      "num_epochs": 3,
      "batch_size": 4,
      "max_seq_length": 1024
    }
  }'

{
  "job_id": "e5f6a7b8-...",
  "dataset_id": "a1b2c3d4-...",
  "base_model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
  "method": "lora",
  "status": "queued",
  "hyperparams": { "lora_r": 16, "lora_alpha": 32, "..." : "..." },
  "progress": 0.0,
  "metrics": {},
  "error": null,
  "created_at": "2026-03-01T12:05:00+00:00",
  "completed_at": null,
  "adapter_path": null
}

# Get a specific job
curl http://localhost:8082/api/v1/training/finetune/{job_id}
 
# List all jobs
curl http://localhost:8082/api/v1/training/finetune

{
  "job_id": "e5f6a7b8-...",
  "status": "completed",
  "progress": 1.0,
  "metrics": {
    "train_loss": 0.42,
    "train_runtime": 1834.5,
    "train_samples_per_second": 12.3
  },
  "adapter_path": "output/lora-sft/final_adapter",
  "completed_at": "2026-03-01T12:35:00+00:00"
}

from isa_model.training.lightning.algorithms.lora_sft_algorithm import LoRASFTAlgorithm
 
algo = LoRASFTAlgorithm(
    model_path="meta-llama/Llama-2-7b-hf",
    lora_r=16,
    lora_alpha=32,
    num_epochs=3,
    learning_rate=2e-4,
)
 
result = algo.create(dataset_path="train.jsonl")
 
print(f"Adapter saved to: {result.adapter_path}")
print(f"Training loss: {result.metrics['train_loss']}")
print(f"Samples: {result.num_samples}, Epochs: {result.num_epochs}")

base-model:adapter-name        # resolves to latest version
base-model:adapter-name:vN     # resolves to specific version

Qwen/Qwen2.5-Coder-1.5B-Instruct:sentiment       # latest version
Qwen/Qwen2.5-Coder-1.5B-Instruct:sentiment:v1    # version 1
meta-llama/Llama-2-7b-hf:code-assistant:v3         # version 3

from isa_model.serving.api.adapter_registry import AdapterRegistry
 
registry = AdapterRegistry()
 
# Register an adapter from a completed job
entry = await registry.register(
    base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
    adapter_name="sentiment",
    adapter_path="output/lora-sft/final_adapter",
    job_id="e5f6a7b8-...",
)
print(entry.model_id)  # "Qwen/Qwen2.5-Coder-1.5B-Instruct:sentiment:v1"
 
# Register another version (auto-increments)
entry_v2 = await registry.register(
    base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
    adapter_name="sentiment",
    adapter_path="output/lora-sft-v2/final_adapter",
    job_id="f6a7b8c9-...",
)
print(entry_v2.model_id)  # "Qwen/Qwen2.5-Coder-1.5B-Instruct:sentiment:v2"
 
# Resolve a model_id to an adapter
adapter = await registry.resolve("Qwen/Qwen2.5-Coder-1.5B-Instruct:sentiment")
print(adapter.adapter_path)  # latest version
 
# List all adapters (latest versions)
all_adapters = await registry.list_adapters()
 
# Filter by base model
qwen_adapters = await registry.list_adapters(
    base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct"
)

1. Merge     LoRA adapter weights into base model
2. Convert   merged model to GGUF format (quantized)
3. Push      GGUF model to Ollama via /api/create

from isa_model.training.export import ModelExporter
 
exporter = ModelExporter(
    ollama_host="http://localhost:11434",
    quantization="q4_0",            # Quantization level for GGUF
)
 
result = await exporter.deploy(
    base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
    adapter_path="output/lora-sft/final_adapter",
    ollama_tag="qwen2-sentiment-v1",
    system_prompt="You are a sentiment analysis assistant.",
)
 
if result.success:
    print(f"Deployed to Ollama as: {result.ollama_tag}")
    print(f"GGUF path: {result.gguf_path}")
else:
    print(f"Deploy failed: {result.error}")

# Direct Ollama usage
ollama run qwen2-sentiment-v1
 
# Via isA Model API
curl -X POST http://localhost:8082/api/v1/invoke \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2-sentiment-v1",
    "provider": "ollama",
    "messages": [{"role": "user", "content": "I absolutely love this product!"}]
  }'

tag = exporter.build_ollama_tag(
    base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
    adapter_name="sentiment",
    version=1,
)
# Returns: "qwen2-5-coder-1-5b-instruct-sentiment-v1"

import asyncio
from isa_model.serving.api.dataset_manager import DatasetManager
from isa_model.serving.api.finetune_manager import FineTuneJobManager
from isa_model.serving.api.adapter_registry import AdapterRegistry
from isa_model.training.export import ModelExporter
 
async def finetune_and_deploy():
    dataset_mgr = DatasetManager()
    job_mgr = FineTuneJobManager()
    registry = AdapterRegistry()
    exporter = ModelExporter()
 
    # 1. Upload dataset
    jsonl = (
        '{"instruction": "Classify sentiment", "input": "Great product!", "output": "positive"}\n'
        '{"instruction": "Classify sentiment", "input": "Terrible service.", "output": "negative"}\n'
    )
    dataset = await dataset_mgr.upload_dataset("sentiment.jsonl", jsonl)
    print(f"Dataset uploaded: {dataset.dataset_id}")
 
    # 2. Launch training job
    job = await job_mgr.create_job(
        dataset_id=dataset.dataset_id,
        base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
        method="lora",
        hyperparams={"lora_r": 16, "num_epochs": 3},
    )
    print(f"Job created: {job.job_id} (status={job.status})")
 
    # 3. (Training runs async — poll for completion)
    # job = await job_mgr.get_job(job.job_id)
 
    # 4. Register adapter checkpoint
    entry = await registry.register(
        base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
        adapter_name="sentiment",
        adapter_path="output/lora-sft/final_adapter",
        job_id=job.job_id,
    )
    print(f"Adapter registered: {entry.model_id}")
 
    # 5. Deploy to Ollama
    result = await exporter.deploy(
        base_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
        adapter_path=entry.adapter_path,
        ollama_tag="qwen2-sentiment-v1",
        system_prompt="You are a sentiment analysis assistant.",
    )
    print(f"Deployed: {result.ollama_tag}")
 
asyncio.run(finetune_and_deploy())

python -c "import torch; print(torch.cuda.is_available())"

pip install peft trl transformers datasets

pip install llama-cpp-python

curl http://localhost:11434/api/tags