charcot-ai/gym-sft

1,801 text and 979 multimodal trajectories on synthetic Charcot Gym tasks; every trajectory scored 1.0.

What it is

This dataset contains trajectories for supervised fine-tuning on synthetic Charcot Gym tasks. Oracle rows were produced by the reference oracle. Rejection-sampling (rs) rows were produced by z-ai/glm-5.3 (open weights, MIT license) on newly generated tasks. Every row was scored 1.0 by the deterministic checker. The rs rows exist only in the text config.

Quick start

Load a text trajectory and format it with a tokenizer compatible with your model:

import json
from datasets import load_dataset

row = load_dataset(
    "charcot-ai/gym-sft", "text", split="train",
    revision="v0.3.0"
)[0]
messages = json.loads(row["messages"])
tools = json.loads(row["tools"])
prompt = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False)

mm_row = load_dataset(
    "charcot-ai/gym-sft", "multimodal", split="train",
    revision="v0.3.0"
)[0]
mm_messages = json.loads(mm_row["messages"])
mm_tools = json.loads(mm_row["tools"])
images = mm_row["images"]
mm_prompt = tokenizer.apply_chat_template(
    mm_messages, tools=mm_tools, tokenize=False
)

Multimodal image bytes are stored in images in message order.

Text composition

Family	Difficulty	Oracle (public)	Oracle (extra)	RS	Total
claims	standard	34	144	126	304
claims	hard	17	72	186	275
clinical_coding	standard	34	144	125	303
clinical_coding	hard	17	72	250	339
document_fhir	standard	34	144	126	304
document_fhir	hard	17	72	187	276
Total		153	648	1000	1801

Multimodal composition

Unchanged from v0.2.0; all rows are oracle rows.

Config	Task source	Family	Difficulty	Trajectories
multimodal	extra	claims	hard	144
multimodal	extra	claims	standard	144
multimodal	extra	clinical_coding	hard	72
multimodal	extra	clinical_coding	standard	144
multimodal	extra	document_fhir	hard	144
multimodal	extra	document_fhir	standard	144
multimodal	public	claims	hard	34
multimodal	public	claims	standard	34
multimodal	public	clinical_coding	hard	17
multimodal	public	clinical_coding	standard	34
multimodal	public	document_fhir	hard	34
multimodal	public	document_fhir	standard	34
Total				979

Schema

Column	Type	Description
task_id	string	Task identifier from the source.
source	string	Producer of the trajectory: oracle or rs (text config). In the multimodal config: public or extra, as in v0.2.0.
task_source	string	Task origin: public, extra or rs; only in the text config.
family	string	Task family.
difficulty	string	Task difficulty.
generator_seed	int64	Generator seed for extra and rs tasks; null for public tasks.
messages	JSON string	Conversation messages used for the trajectory.
tools	JSON string	Tool definitions available to the model.
images	list	PNG images in message order; only in the multimodal config.

How it was made

The oracle runs each task through the official runner; only episodes scored 1.0 are kept. Public rows come from public tasks. Extra rows use the public generators with the eight seeds in seeds.json and are deduplicated by canonical task content.

The rs rows come from rejection sampling. The public generators produced new tasks from a separate set of seeds; these tasks were deduplicated and checked against all other task sets. z-ai/glm-5.3 solved each task through the official runner at temperature 0.7. A row was accepted only when the deterministic checker scored it 1.0.

For every Hard task and for Clinical coding Standard, the teacher prompt also included a plan derived from the oracle. The plan and the system prompt are not in the stored rows. Rows whose assistant messages refer to the plan were rejected.

Held-out tasks

The 36 public task IDs in heldout.json are excluded from training and reserved for evaluation. The generated held-out v2 evaluation tasks are excluded too.

Intended use and limits

Use these trajectories for research on supervised fine-tuning and evaluation. They are synthetic and not for clinical use.

Contamination

Every trajectory comes from a public-split task or from the public generators. The rs tasks share no ID, seed, prompt, initial state or canonical content with the public split, held-out v1 or v2, the extra rows, the RL tasks or the private split. No private-split task, seed or trajectory is used.

License and citation

Charcot Gym is licensed under Apache-2.0.

@software{charcot_gym,
  title = {Charcot Gym},
  author = {Charcot AI},
  version = {0.1.0},
  url = {https://github.com/Charcot-AI/gym}
}

Related

Charcot Gym