---
library_name: transformers
tags:
- text-generation
- llama-3.1
- computational-social-science
- user-simulation
- conditioned-comment-prediction
- social-media
base_model: meta-llama/Llama-3.1-8B-Instruct
license: llama3.1
language:
- en
---

# Model Card for echo-Llama-3.1-8B-Instruct

This model performs Conditioned Comment Prediction (CCP), designed to act as a "silicon subject" for computational social science. It predicts how a specific social media user will respond to a given stimulus by utilizing both an explicit user profile and implicit behavioral history.

## Model Details

### Model Description

The model was fine-tuned using Supervised Fine-Tuning (SFT) to optimize its ability to generate high-fidelity, user-specific replies to online content. It isolates the capability of response generation, prioritizing operational validity over surface-level plausibility by benchmarking against authentic digital traces. 

- **Developed by:** Nils Schwager, Simon Münker, Alistair Plum, Achim Rettinger (Trier University, University of Luxembourg)
- **Funded by:** EU's Horizon Europe Framework (HORIZON-CL2-2022-DEMOCRACY-01-07)
- **Model type:** Autoregressive Large Language Model (Instruction-tuned)
- **Language(s) (NLP):** English (en)
- **License:** llama3.1
- **Finetuned from model:** meta-llama/Llama-3.1-8B-Instruct

### Model Sources

- **Repository:** https://github.com/nsschw/Conditioned-Comment-Prediction
- **Paper:** Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction (arXiv:2602.22752v1)

## Uses

### Direct Use

The model is built for computational social scientists and researchers modeling discourse dynamics and individual-level behavioral patterns. It is strictly optimized for predicting first-order text replies to textual stimuli (such as social media posts or news articles) given a specific user context.

### Out-of-Scope Use

The model cannot process multi-modal inputs (e.g., URLs, images, GIFs) or simulate non-verbal interactions (e.g., liking behavior). It is not intended for generating coordinated inauthentic behavior, micro-targeted disinformation, or impersonating individuals for deceptive purposes.

## Bias, Risks, and Limitations

- **Privacy:** The training utilized authentic, public digital traces from real X (Twitter) users. These individuals did not provide explicit informed consent for their communication patterns to be replicated by generative models.
- **Dual-Use Risks:** The framework demonstrated here enables the generation of synthetic content that mimics individual communication patterns with measurable fidelity. Malicious actors could exploit this for sophisticated bot campaigns or targeted persuasion.
- **Semantic Boundaries:** While SFT aligns the surface-level structure of the text output, the model's semantic grounding remains bounded by the representational capacity of the 8B parameter class.

## How to Get Started with the Model

The model requires a combined explicit and implicit conditioning format. Structure your prompt as a native chat sequence containing the generated user profile (system prompt) followed by up to 29 historical interactions. 

```text
System: [Insert Generated Biography]
User: Comment on the following content: [Historical Stimulus 1]
Assistant: [Authentic Historical Reply 1]
...
User: Comment on the following content: [New Target Stimulus]
Assistant:
```

## Training Details

### Training Data
The model was fine-tuned on a corpus of English X (Twitter) discourse collected up to August 2023.

* **Sample Size:** 3,800 politically active users.
* **Context Limit:** Up to 30 stimulus-response interactions per user.
* **Exclusions:** Interactions containing media, and users with fewer than four historical replies.

### Training Procedure

#### Training Hyperparameters
* **Training regime:** 8-bit quantization, Paged AdamW optimizer.
* **Epochs:** 1
* **Max sequence length:** 4,500 tokens
* **Training paradigm:** SFT on complete input sequences (system prompt, user prompts, and model completions) using TRL defaults.

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data
Evaluations were conducted on a deterministic, held-out test split of 650 English users to prevent cross-user data leakage.

#### Metrics
The evaluation utilized automated metrics across five independent generation runs (decoding temperature 0.75, max new tokens 500) to capture both lexical overlap and semantic alignment.

* **BLEU & ROUGE-1:** Measures precision-oriented n-gram overlap and unigram overlap.
* **Length Ratio:** Quantifies output volume alignment against the authentic reference length.
* **Embedding Distance:** Assesses semantic intent alignment via cosine distance using Qwen3-Embedding-8B.

### Results
* **BLEU:** 0.083
* **ROUGE-1:** 0.229
* **Length Ratio:** 0.961
* **Embedding Distance:** 0.397

> **Observation:** Explicit biography conditioning becomes largely redundant post-fine-tuning, as the model successfully performs latent inference directly from the provided behavioral histories.

## Technical Specifications

### Compute Infrastructure

#### Hardware
The model was fine-tuned on a single NVIDIA L40S GPU (48GB VRAM).

## Citation

**BibTeX:**
```bibtex
@article{schwager2026towards,
  title={Towards Simulating Social Media Users with LLMs: Evaluating the Operational Validity of Conditioned Comment Prediction},
  author={Schwager, Nils and M{\"u}nker, Simon and Plum, Alistair and Rettinger, Achim},
  journal={arXiv preprint arXiv:2602.22752},
  year={2026}
}