slipstream-finetune

name	slipstream-finetune
description	Finetune LLMs to speak Slipstream natively - complete guide with GLM-4-9B

Slipstream Finetuning Guide

Train LLMs to communicate using the Slipstream protocol natively. This guide covers dataset generation, model finetuning, and releasing on HuggingFace.

Recommended Model: GLM-4-9B-0414

Why GLM-4-9B-0414?

• MIT licensed (can release finetuned weights commercially)
• 9B parameters - good balance of capability and trainability
• Specifically optimized for function calling and agentic tasks
• Excellent instruction following

Quick Start

1. Generate High-Quality Dataset

Option A: Template-based (fast, free)

python -m slipcore.finetune -n 1000 -f sharegpt -o slipstream_train.jsonl

Option B: LLM-enhanced (higher quality, requires API)

# Using Claude API (recommended for quality)
export ANTHROPIC_API_KEY="your-key"
python -m slipcore.finetune_llm -n 1000 --provider anthropic -o slipstream_train.jsonl

# Using OpenAI (good quality, widely available)
export OPENAI_API_KEY="your-key"
python -m slipcore.finetune_llm -n 1000 --provider openai --model gpt-4o-mini -o slipstream_train.jsonl

# Using Together.ai (cheaper, good for large datasets)
export TOGETHER_API_KEY="your-key"
python -m slipcore.finetune_llm -n 2000 --provider together --model meta-llama/Llama-3.3-70B-Instruct-Turbo -o slipstream_train.jsonl

# Using DeepSeek (very cheap, good quality)
export DEEPSEEK_API_KEY="your-key"
python -m slipcore.finetune_llm -n 2000 --provider deepseek -o slipstream_train.jsonl

Cost Estimates for 1000 examples:

Provider	Model	~Cost
Anthropic	claude-sonnet-4-20250514	~$0.50
OpenAI	gpt-4o-mini	~$0.15
Together	Llama-3.3-70B	~$0.10
DeepSeek	deepseek-chat	~$0.02

2. Finetune GLM-4-9B with Unsloth

from unsloth import FastLanguageModel
import torch

# Load GLM-4-9B-0414
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="THUDM/GLM-4-9B-0414",
    max_seq_length=2048,
    dtype=None,  # Auto-detect
    load_in_4bit=True,  # QLoRA - fits in ~8GB VRAM
)

# Add LoRA adapters
model = FastLanguageModel.get_peft_model(
    model,
    r=16,  # LoRA rank
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=42,
)

# Load dataset
from datasets import load_dataset
dataset = load_dataset("json", data_files="slipstream_train.jsonl", split="train")

# GLM-4 chat template
def format_glm4(example):
    convs = example["conversations"]
    text = ""
    for conv in convs:
        if conv["from"] == "system":
            text += f"[gMASK]<sop><|system|>\n{conv['value']}"
        elif conv["from"] == "human":
            text += f"<|user|>\n{conv['value']}"
        elif conv["from"] == "gpt":
            text += f"<|assistant|>\n{conv['value']}"
    return {"text": text}

dataset = dataset.map(format_glm4)

# Train
from trl import SFTTrainer
from transformers import TrainingArguments

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=2048,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=10,
        max_steps=200,  # ~1000 examples, 2 epochs
        learning_rate=2e-4,
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        logging_steps=10,
        output_dir="slipstream_glm4",
        optim="adamw_8bit",
        seed=42,
    ),
)

trainer.train()

# Save LoRA adapter
model.save_pretrained("slipstream_glm4_lora")
tokenizer.save_pretrained("slipstream_glm4_lora")

3. Test the Finetuned Model

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="slipstream_glm4_lora",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

# Test
prompt = """[gMASK]<sop><|system|>
You communicate using the Slipstream protocol.<|user|>
Tell the backend team to review the authentication code<|assistant|>
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0]))
# Expected: SLIP v1 agent backend RequestReview auth_code

4. Export and Release

Option A: LoRA adapter only (~200MB)

model.push_to_hub("your-username/slipstream-glm4-9b-lora")
tokenizer.push_to_hub("your-username/slipstream-glm4-9b-lora")

Option B: Merged full model (~18GB)

# Merge LoRA into base model
merged_model = model.merge_and_unload()
merged_model.push_to_hub("your-username/slipstream-glm4-9b")
tokenizer.push_to_hub("your-username/slipstream-glm4-9b")

Option C: GGUF for Ollama/llama.cpp (~5-9GB)

# Save as GGUF quantized
model.save_pretrained_gguf(
    "slipstream_glm4_gguf",
    tokenizer,
    quantization_method="q4_k_m",  # Good balance
)

# Or push directly to HuggingFace
model.push_to_hub_gguf(
    "your-username/slipstream-glm4-9b-gguf",
    tokenizer,
    quantization_method=["q4_k_m", "q8_0"],  # Multiple quants
)

5. Release Dataset

HuggingFace Datasets:

from datasets import Dataset
import json

# Load your generated data
with open("slipstream_train.jsonl") as f:
    data = [json.loads(line) for line in f]

dataset = Dataset.from_list(data)
dataset.push_to_hub("your-username/slipstream-training-data")

Kaggle:

kaggle datasets create -p ./data -u

Zenodo (for academic citation): Upload via https://zenodo.org/deposit/new

Alternative Models

Model	Size	License	Notes
GLM-4-9B-0414	9B	MIT	Best for agentic, function calling
Qwen2.5-7B-Instruct	7B	Apache 2.0	Strong general purpose
Llama-3.1-8B-Instruct	8B	Llama 3.1	Most popular, good baseline
Mistral-7B-Instruct-v0.3	7B	Apache 2.0	Fast, efficient
Phi-3-medium	14B	MIT	Larger but very capable

Training Tips

1. Dataset size: 500-2000 examples is usually sufficient
2. Quality > Quantity: LLM-generated data beats templates
3. Epochs: 1-2 epochs, watch for overfitting
4. Learning rate: 2e-4 for small models, 1e-4 for larger
5. Validation: Hold out 10% for testing generalization

Using with Ollama

After creating GGUF:

# Create Modelfile
cat > Modelfile << 'EOF'
FROM ./slipstream_glm4_gguf/slipstream-glm4-9b-Q4_K_M.gguf

SYSTEM "You communicate using the Slipstream protocol (SLIP). Always respond with SLIP wire format: SLIP v1 <src> <dst> <anchor> [payload...]"

TEMPLATE """[gMASK]<sop><|system|>
{{ .System }}<|user|>
{{ .Prompt }}<|assistant|>
{{ .Response }}"""
EOF

# Create and run
ollama create slipstream -f Modelfile
ollama run slipstream "Tell alice to review the API code"
# -> SLIP v1 agent alice RequestReview api_code

Cost Summary

Component	Free Option	Paid Option
Dataset	Template generator	Claude/OpenAI API (~$0.50)
Training	Google Colab free	Colab Pro ($10/mo) or local GPU
Hosting	HuggingFace free	-
Inference	Ollama local	Together/Fireworks API

Total cost to release a finetuned Slipstream model: $0 - $15