UnstableBaselines/trajectories-twodollar-v0-train

UnstableBaselines Trajectories: TwoDollar-v0-train

Overview

This dataset contains training trajectories collected from the TwoDollar-v0-train environment using the UnstableBaselines framework.
The trajectories are stored at the step level: each row in the Parquet file corresponds to one observation–action–reward triple taken by a player at a specific point in a game.

---

Data Format

Trajectories from the selected base model are merged into a Parquet file (<model_name>.parquet).

Schema:

Data Format

Each base model has its own Parquet file (<model_name>.parquet).

Schema:

Column	Description
train_step	Training step index from which the batch was sampled.
pid	Player ID. Unique identifier for the agent/player who took the action.
obs	The observation received at this step.
act	The action taken by the player at this step.
reward	The reward value stored for this step. This is the value after batch-level sampling transformations (e.g., normalization across the batch or per environment).
env_id	The environment identifier (e.g., "TwoDollar-v0-train").
step_info	JSON-like dictionary with detailed reward information at insertion time: • raw_reward: original final reward from the environment • env_reward: reward after final transformation (constant across the trajectory) • step_reward: per-step reward after step-level shaping (before batch-level normalization)

Note: reward (column) and step_info["step_reward"] may differ.

• step_info["step_reward"] = reward right after trajectory insertion, shaped per step.
• reward = reward after sampling transformations, such as NormalizeRewardsByEnv, applied to the batch before it is written out.

---

Configurations by Model

Qwen/Qwen2.5-7B-Instruct

• LoRA settings:
  • Rank = 16, Alpha = 32, Dropout = 0.0
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• vLLM config:
  • Temperature = 0.6
  • Max tokens = 4096
  • Max model len = 8192
  • Max parallel seq = 128
  • Max loras = 8
• Reward transformations:
  • Final reward: RoleAdvantageByEnvFormatter
  • Step reward: RewardForFormat(1.5), PenaltyForInvalidMove(1.0, -1.0)
  • Sampling reward: NormalizeRewardsByEnv(True)
• Training setup:
  • Learner: REINFORCELearner
  • Batch size = 384, Mini-batch size = 1, LR = 1e-5, Grad clip = 0.2
  • Gradient + activation checkpointing: enabled
  • Collector: 384 train workers, 16 eval workers

---

Usage

You can load the dataset with:

from datasets import load_dataset

# Load entire dataset
ds = load_dataset("UnstableBaselines/trajectories-twodollar-v0-train", split="train")

# Or load a specific model parquet (if multiple are available)
ds_qwen = load_dataset("UnstableBaselines/trajectories-twodollar-v0-train", data_files="qwen257binstruct.parquet")["train"]