Instructions to use 294-288-pretrain-interaction/160M-rtp with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 294-288-pretrain-interaction/160M-rtp with Transformers:

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

pipe = pipeline("text-generation", model="294-288-pretrain-interaction/160M-rtp")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("294-288-pretrain-interaction/160M-rtp", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use 294-288-pretrain-interaction/160M-rtp with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "294-288-pretrain-interaction/160M-rtp"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "294-288-pretrain-interaction/160M-rtp",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/294-288-pretrain-interaction/160M-rtp

SGLang

How to use 294-288-pretrain-interaction/160M-rtp 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 "294-288-pretrain-interaction/160M-rtp" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "294-288-pretrain-interaction/160M-rtp",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "294-288-pretrain-interaction/160M-rtp" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "294-288-pretrain-interaction/160M-rtp",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use 294-288-pretrain-interaction/160M-rtp with Docker Model Runner:
```
docker model run hf.co/294-288-pretrain-interaction/160M-rtp
```

160M RTP dose campaign

This repository collects Yihan's 24 research runs: six injection doses (10, 25, 50, 100, 250, 500) and four source allocations (M, C, G, MCG). The exact trigger is dAtA[iNdEx]. Models use the Pythia-160M architecture and random initialization with seed 42. Each run targets 16,000,090,112 tokens on the frozen Dolma Math/Code/General mixture.

The campaign was interrupted on 2026-10-13 UTC after 23 of 24 runs were released. All six doses of M, C and G, and doses 10 through 250 of MCG, have complete checkpoint and evaluation releases: 391 full training checkpoints and 368 integer-B evaluation stages. releases/<run>.json records uploaded file checksums and the verified Hub revision. The repository is public, matching the group's reference release.

MCG500 remains incomplete. Its seven scheduled checkpoints at 1B, 2B, 3B, 3.2B, 4B, 5B and 6B were backed up before compute became unavailable. The latest saved state is runs/MCG500/step-45777.pt (6,000,082,944 tokens). These backups do not constitute a completed run, and MCG500 has no completed evaluation release. Only a verified releases/<run>.json counts as a complete run; missing results are pending, not zero-valued measurements.

See the result tables and 1B-16B trajectory figures for the 23 verified experiments. The pending MCG500 panels are explicitly blank.

Layout

models/<run>/: final safetensors weights, config and tokenizer, directly loadable.
runs/<run>/: completed runs have 17 full training states (1B through 16B, plus 3.2B), training status, timing, evaluation outputs and verification receipts. The incomplete MCG500 directory contains only its seven backed-up states.
evaluations/<scheme>/dose-<dose>/Clean/: matching Clean evaluation outputs.
releases/: per-run upload and verification records.

The public release contains models, checkpoints and evaluation results. Model folders include the config and tokenizer required to load the weights. Frozen plans, manifests, poison payloads, training/evaluation configuration files and source archives remain in the private working archive. Result receipts retain input hashes for integrity checks without publishing those input files.

Example loading the released M10 model:

from transformers import AutoTokenizer, GPTNeoXForCausalLM
repo = "294-288-pretrain-interaction/160M-rtp"
model = GPTNeoXForCausalLM.from_pretrained(repo, subfolder="models/M10")
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder="models/M10")

The .pt files contain model, optimizer, scheduler, all eight ranks' RNG states, step, token count, injection counters and manifest/config hashes. They are full training states, not from_pretrained directories. Smoke checkpoints and unfinished .part files are excluded from releases.

Measurements

Each completed run evaluates all 16 integer-B checkpoints. The primary RTP effect is the change caused by the trigger in the sequence-log-probability preference for a toxic versus detoxified continuation, minus the same contrast in Clean. Three pairs per Math/Code/General prompt are averaged before computing paired prompt SEM; native RTP contexts are also reported.

Performance is natural validation NLL minus Clean, separately by domain. Additional diagnostics include five sampled continuations per held-out prompt and trigger condition scored by the pinned Pythia-1.4B model, plus the 5,153-example zero-shot LAMBADA evaluation. SEM describes prompt/example variation; there is only one training seed, not repeated-seed uncertainty.

Clean checkpoints come from the group's earlier A100 training trajectory; the 24 poison trajectories run on H100. Cross-hardware numerical differences are an additional limitation of comparisons against this reused baseline.

The plan records trigger_confirmed_event_selection_draft: the trigger was confirmed, and nested source-stratified event subsets follow the existing dose sweep. This label does not claim separate group approval of the event-selection protocol. The full input definitions are retained in the private working archive; released checkpoint and evaluation records include actual injection counts.

These are experimental research models trained with controlled toxic payloads.

Downloads last month: -; Downloads are not tracked for this model. How to track