Instructions to use emergent-wisdom/thl-llama-3.3-70b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use emergent-wisdom/thl-llama-3.3-70b-lora with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("unsloth/llama-3.3-70b-instruct-bnb-4bit")
model = PeftModel.from_pretrained(base_model, "emergent-wisdom/thl-llama-3.3-70b-lora")

Transformers

How to use emergent-wisdom/thl-llama-3.3-70b-lora with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="emergent-wisdom/thl-llama-3.3-70b-lora")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("emergent-wisdom/thl-llama-3.3-70b-lora", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use emergent-wisdom/thl-llama-3.3-70b-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "emergent-wisdom/thl-llama-3.3-70b-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "emergent-wisdom/thl-llama-3.3-70b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/emergent-wisdom/thl-llama-3.3-70b-lora

SGLang

How to use emergent-wisdom/thl-llama-3.3-70b-lora with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "emergent-wisdom/thl-llama-3.3-70b-lora" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "emergent-wisdom/thl-llama-3.3-70b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "emergent-wisdom/thl-llama-3.3-70b-lora" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "emergent-wisdom/thl-llama-3.3-70b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Studio

How to use emergent-wisdom/thl-llama-3.3-70b-lora with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for emergent-wisdom/thl-llama-3.3-70b-lora to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for emergent-wisdom/thl-llama-3.3-70b-lora to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for emergent-wisdom/thl-llama-3.3-70b-lora to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="emergent-wisdom/thl-llama-3.3-70b-lora",
    max_seq_length=2048,
)

Docker Model Runner
How to use emergent-wisdom/thl-llama-3.3-70b-lora with Docker Model Runner:
```
docker model run hf.co/emergent-wisdom/thl-llama-3.3-70b-lora
```

THL-Llama-3.3-70B-LoRA

A LoRA adapter for Llama-3.3-70B-Instruct trained using Temporal Hindsight Learning (THL) — a framework that converts raw historical logs into high-quality reasoning supervision by having a Teacher model generate "ideal prediction" traces using hindsight.

Paper: Temporal Hindsight Learning: Training Calibrated Reasoners with Future Oracles

Key Result

The THL Student (this model) achieves accuracy parity with its Teacher (Gemini 3 Flash, a ~1T frontier model) on genuinely unseen February–May 2025 events (Mann-Whitney U, p=0.97), while significantly outperforming the untrained base model (p<0.001). A 70B model trained on 505 reasoning traces matches a model with an order of magnitude more parameters.

Model Details

Base model: meta-llama/Llama-3.3-70B-Instruct (4-bit quantized via Unsloth)
Adapter: LoRA (rank 16, alpha 16) targeting all linear modules (q, k, v, o, gate, up, down projections)
Quantization: 4-bit NF4 via QLoRA
Knowledge cutoff: December 2023 (base model)
Developed by: Henrik Westerberg / Emergent Wisdom
License: MIT

Training

Method: Supervised fine-tuning on Teacher-generated "Ideal Prediction" reasoning traces
Teacher: Gemini 3 Flash Preview (knowledge cutoff: January 2025)
Training data: 505 reasoning traces across 106 global events from January–December 2024
Reasoning structure: Each trace follows the Forecasting Pentagon — five angles (Structural/Mechanism, Economic/Incentives, Political/Social, Base Rates/Precedents, Temporal/Pacing)
Hardware: NVIDIA T4 on Google Colab Free Tier
Framework: Unsloth + PEFT 0.18.1

Evaluation

Evaluated on 15 genuinely unseen events from February–May 2025 (75 exam prompts across all five Pentagon angles). Scored by Claude Opus 4.6 (independent Auditor from a different model family than the Teacher).

Model	Role	Median Score	vs Base
Base Llama-3.3	Frozen (Dec 2023)	Lower	—
THL Student	Trained on 2024	Higher	p<0.001
Gemini 3 Flash	Teacher baseline	Comparable	p=0.97 vs THL

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.3-70B-Instruct",
    load_in_4bit=True,
)
model = PeftModel.from_pretrained(base_model, "emergent-wisdom/thl-llama-3.3-70b-lora")
tokenizer = AutoTokenizer.from_pretrained("emergent-wisdom/thl-llama-3.3-70b-lora")

Reasoning Format

The model produces structured reasoning traces:

**What I know from the context:**
[Analysis of provided context clues]

**What I know from my training (pre-2024):**
[Relevant pre-cutoff knowledge]

**Causal Analysis:**
[Step-by-step causal reasoning]

**My Prediction:**
[Calibrated prediction with probability estimates]

Citation

@misc{westerberg2026thl,
  author = {Westerberg, Henrik},
  title = {Temporal Hindsight Learning: Training Calibrated Reasoners with Future Oracles},
  year = {2026},
  publisher = {Emergent Wisdom}
}

Repository

Code & data: emergent-wisdom/temporal-hindsight-learning
Website: emergentwisdom.org

Downloads last month: 2