Skip to content
Other

actionproof-proc-a12-state-reward-s1

LoRA adapter for database agent action verification

2 min read

This research LoRA adapter (a lightweight add-on on top of a base model) was created as part of the ActionProof experiment to train AI agents in the ProcureSim procurement simulator. In this environment, the system response to a data write looks identical even if the operation partially executed, duplicated, or failed completely, forcing the agent to verify the actual database state. Trained on top of CohereLabs/tiny-aya-global with a warm start by author atharvgaur18, the adapter uses GRPO (reinforcement learning with group-based response evaluation) and a binary reward tied solely to confirmed database states. It is useful for AI agent researchers studying tool-call reliability and protection against "phantom" operations.

Key features

  • Binary database-state reward: during GRPO training, the model receives 1 only when every requested step is confirmed by the verifier directly in the database.
  • Protection against false confirmations: the adapter was trained under conditions where the write-success message is identical across full, partial, duplicated, or zero execution.
  • Two-stage weight assembly: a fresh LoRA (r = 16) is applied on top of the CohereLabs/tiny-aya-global base with the pre-merged actionproof-proc-warm-start adapter.
  • Experiment reproducibility: run parameters are locked in run_config.json, and all metrics tie back to report files and preregistration in the code repository.

What you can do

  • Reproduce the ActionProof experiment: run agent evaluations in the ProcureSim environment from the author's repository and compare behavior under unreliable API responses.
  • Study agent self-verification patterns: analyze how the model performs state read-backs after write operations in the simulated database.
  • Compare different seeds and rewards: test this variant (seed 1) alongside neighboring checkpoints in the s0–s4 series to evaluate training stability.

Example

from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="bfloat16")
base = PeftModel.from_pretrained(base, "atharvgaur18/actionproof-proc-warm-start").merge_and_unload()
model = PeftModel.from_pretrained(base, "atharvgaur18/actionproof-proc-a12-state-reward-s1")
Read moreDetails
Metrics on proc_conf (n = 300, greedy): verified safe completion — 0.793, phantom read-back — 253/253.
Training: TRL 1.7.1 GRPO via AgentTune, 200 steps, 4 prompts × 8 rollouts, lr 2e-5, KL 0.02, T = 1, 16 turns, LoRA r = 16.
Limitations: research artifact strictly for the ProcureSim procurement simulation, not intended for real payments.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 6.4 GB · ready files7.5 GB
  • FP8 / INT8recommendedweights 3.2 GB4.1 GB
  • 4-bitweights 1.9 GB2.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 5060 Ti 8GB

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model tiny-aya-global.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # atharvgaur18/actionproof-proc-a12-state-reward-s1
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download CohereLabs/tiny-aya-global --local-dir ~/models/tiny-aya-global
    hf download atharvgaur18/actionproof-proc-a12-state-reward-s1 --local-dir ~/models/actionproof-proc-a12-state-reward-s1
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to tiny-aya-global
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

CohereLabs/tiny-aya-global

This model is a LoRA / adapter. Download the base model first, then load this one on top.

  1. tiny-aya-basefamily root
  2. tiny-aya-global
  3. actionproof-proc-a12-state-reward-s1

New posts about versions of tiny-aya-base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face

About the model

atharvgaur18/actionproof-proc-a12-state-reward-s1
Author
atharvgaur18
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments