risk-averse-ai-adapter-archive

Risk-Averse AI Adapter Archive

Model artifacts for the paper Out-of-Distribution Generalization of Risk Aversion in Language Models, by Kristina Zhang, Junior Chinomso Okoroafor, Benjamin Maltbie, Andrew Lin, Abhitej Bokka, and Elliott Thornley (Massachusetts Institute of Technology).

This archive holds only the trained weights and the steering directions, because they are large binaries. The code, the datasets, the construction inputs, and the result tables are in the paper's code repository:

https://github.com/riskaverseAIs/riskaverseAIs

Contents

paper_adapters/
  sft/           <model>/seed-{1..5}/   LoRA adapters
  tie_training/  <model>/seed-{1..5}/   LoRA adapters
  dpo/           <model>/seed-{1..5}/   LoRA adapters
  rmft/          <model>/seed-{1..5}/   LoRA adapters and reward heads
  steering/      <model>/seed-{1..5}/   direction vectors
manifests/
  adapter_manifest.json

Five base models, five methods, five seeds each:

Every artifact here backs a number in the paper. There is nothing else in the archive.

Which artifact backs which result

manifests/adapter_manifest.json has one entry per fine-tuned adapter, giving the method, base model, seed, path, LoRA geometry, and a content hash. The steering directions carry their own manifest at paper_adapters/steering/steering_manifest.json, which additionally records the steered layer, the raw multiplier alpha, the mean residual-stream norm at that layer, and the ratio strength_r between them.

Using an adapter

Each fine-tuning artifact is a PEFT LoRA directory: load the base model, then apply the adapter.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import snapshot_download

base = "Qwen/Qwen3-8B"
path = snapshot_download(
    "MIT-SERC-risk-averse-AIs/risk-averse-ai-adapter-archive",
    allow_patterns="paper_adapters/sft/qwen3-8b/seed-1/*",
)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, f"{path}/paper_adapters/sft/qwen3-8b/seed-1")
tok = AutoTokenizer.from_pretrained(base)

RMFT adapters wrap a reward-model backbone rather than a causal language-model head; they use PEFT task type FEATURE_EXTRACTION and are loaded together with the reward head stored beside them.

Steering directions are not adapters. Each is a single unit-normalised vector added to the residual stream at one layer, at every token position, during generation. See paper_adapters/steering/README.md.

Shared LoRA configuration

Rank 32, alpha 64, dropout 0.05, no bias, 4-bit quantised base model, targeting q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.

Verification

See PROVENANCE.md.

Citation

@misc{zhang2026oodriskaversion,
  title  = {Out-of-Distribution Generalization of Risk Aversion in Language Models},
  author = {Zhang, Kristina and Okoroafor, Junior Chinomso and Maltbie, Benjamin
            and Lin, Andrew and Bokka, Abhitej and Thornley, Elliott},
  year   = {2026},
  note   = {\url{https://github.com/riskaverseAIs/riskaverseAIs}}
}