Reddit Thread Simulation Workflow

This README documents the current end-to-end pipeline used in this project:

1. Build raw Reddit thread trees with depth filtering
2. Build seed-A-B-C-D chains with metadata
3. Run agent simulation
4. Score generated texts with Detoxify
5. Analyze results in notebook

0) Environment

From project root:

cd /u/yian3/toxic_agent

Install core dependencies if needed:

pip install pandas networkx torch detoxify tqdm jupyter

---

1) Build Reddit trees (depth >= 5)

Script: data/build_reddit.py

This script reconstructs nested trees from raw Reddit branch-style JSON and writes:

• data/extracted/politics_depth_ge5.jsonl

Run:

python /u/yian3/toxic_agent/data/build_reddit.py

Notes:

• In the current script, input/output paths are set inside __main__.
• Default min_depth=5.

---

2) Build chain records for simulation

Script: data/build_threads.py

This extracts one deepest chain per thread and writes:

• seed, A, B, C, D text and metadata (created_utc, score, etc.)
• Detoxify scores for each role
• Output file: data/extracted/politics_seedA_BCD_chains_detoxify.jsonl

Run:

python /u/yian3/toxic_agent/data/build_threads.py

---

3) Run influence simulation

Script: src/run_influence_baseline.py

Run from src/ so -m run_influence_baseline resolves correctly:

cd /u/yian3/toxic_agent/src

python -m run_influence_baseline \
  --seed_source reddit_jsonl \
  --reddit_jsonl /projects/bcxt/agent/data/reddit/extracted/politics_seedA_BCD_chains_detoxify.jsonl \
  --seed_strategy root \
  --reddit_require_max_depth -1 \
  --out_jsonl ../data/influence_baseline_threads_reddit.jsonl \
  --out_summary ../data/influence_baseline_summary_reddit.json \
  --model gpt-4o-mini \
  --n_seeds 50 \
  --compute_sentiment

position 1 python -m run_influence_baseline
--seed_source reddit_jsonl
--reddit_jsonl /projects/bcxt/agent/data/reddit/extracted/politics_seedA_BCD_chains_detoxify.jsonl
--seed_strategy root
--reddit_require_max_depth -1
--out_jsonl ../data/influence_baseline_threads_detoxify_200.jsonl
--out_summary ../data/influence_baseline_summary_detoxify_200.json
--model gpt-4o-mini
--n_seeds 200
--compute_sentiment
--intervention_position pos1

Current simulation behavior:

• Agents can vote and/or reply at each opportunity.
• Vote events immediately update the target message score.
• Message created_utc is tracked for simulated replies.
• Vote events are stored in the output JSONL under votes.

---

4) Add Detoxify scores to generated messages

Script: data/read_jsonl.py

This script:

• loads influence_baseline_threads_reddit.jsonl
• scores each generated message with Detoxify
• writes scores back into each message under message["detoxify"]

Run:

cd /u/yian3/toxic_agent/data
python read_jsonl.py

Note:

• The path is currently set in the script (path = "influence_baseline_threads_reddit.jsonl").

---

5) Analysis in notebook

Notebook: data/read_jsonl.ipynb

Use this notebook for:

• per-mode comparisons (toxic, neutral, removed)
• turn-level sentiment/toxicity trends
• vote behavior and score changes over turns
• summary tables and plots

Run:

cd /u/yian3/toxic_agent/data
jupyter notebook read_jsonl.ipynb

6) Analysis in python code

python -m evaluate_single_thread_influence \
  --mild_path ../data/reddit/influence_baseline_threads_detoxify_mild.jsonl \
  --medium_path ../data/reddit/influence_baseline_threads_detoxify_medium.jsonl \
  --strong_path ../data/reddit/influence_baseline_threads_detoxify_strong.jsonl \
  --analysis_dir ../analysis

---

Outputs You Should See

• data/extracted/politics_depth_ge5.jsonl
• data/extracted/politics_seedA_BCD_chains_detoxify.jsonl
• data/influence_baseline_threads_reddit.jsonl
• data/influence_baseline_summary_reddit.json