agentic-sft / README.md
VanishD's picture
Add files using upload-large-folder tool
b009c0d verified
|
Raw History Blame Contribute Delete
5.18 kB
metadata
license: other
task_categories:
  - text-generation
language:
  - en
tags:
  - agentic
  - sft
  - arc-agi
  - frontier-cs
  - tool-use
  - reasoning
  - trajectories
size_categories:
  - 100K<n<1M

Agentic SFT trajectories (ARC-AGI-2 + frontier-cs-algo)

All corpora under train/data/ as one dataset. Each subdirectory is one build (task × rollout context × selection × training cap). The published single-corpus set VanishD/arc-agi-2-agentic-sft is the arc2_95k_win_128k_old slice of this tree.

Live builds (no _old suffix): 318,497 samples across 6,060 trajectories. A sample is one assistant turn plus the exact context that turn was produced from.

Corpora

corpus cap select samples traj max tokens ctx_trimmed
arc2_48k_all-s0_64k 65536 all 19,766 588 65,531 33
arc2_48k_all-s50_64k 65536 all (floor=0.5) 15,173 547 65,531 16
arc2_48k_best-s0_64k 65536 best 14,626 586 65,501 10
arc2_48k_best-s50_64k 65536 best (floor=0.5) 12,380 547 65,501 6
arc2_95k_win_128k 131072 win 28,408 952 116,528 0
arc2_95k_win_32k 32768 win 28,333 952 32,768 8,141
arc2_95k_win_72k 73728 win 28,405 952 73,726 766
fcs_48k_all-s0_64k 65536 all 23,391 91 65,398 32
fcs_48k_all-s50_64k 65536 all (floor=50.0) 14,884 61 65,370 11
fcs_48k_best-s0_64k 65536 best 5,474 91 65,313 7
fcs_48k_best-s50_64k 65536 best (floor=50.0) 3,486 61 65,313 3
fcs_95k_all-s0_100k 100000 all 26,896 94 100,000 2,021
fcs_95k_all-s0_128k 131072 all 26,899 94 126,997 0
fcs_95k_all-s50_100k 100000 all (floor=50.0) 16,278 59 100,000 1,190
fcs_95k_all-s50_128k 131072 all (floor=50.0) 16,280 59 126,997 0
fcs_95k_best-s0_100k 100000 best 10,088 94 99,999 725
fcs_95k_best-s0_128k 131072 best 10,090 94 126,997 0
fcs_95k_best-s50_100k 100000 best (floor=50.0) 6,298 59 99,999 404
fcs_95k_best-s50_128k 131072 best (floor=50.0) 6,299 59 126,997 0
fcs_95k_eval20_128k 131072 all 5,043 20 114,188 0

_old builds predate require_tool_call and are kept only to reproduce older numbers:

corpus cap select samples traj max tokens ctx_trimmed
arc2_95k_win_128k_old 131072 win 29,358 952 116,528 0
arc2_95k_win_32k_old 32768 win 29,283 952 32,768 8,431
arc2_95k_win_72k_old 73728 win 29,355 952 73,726 806

Naming: <task>_<rollout-ctx>_<selection>_<train-cap>. _ separates the four fields; - joins inside one field. Score floors: ARC is 0–1 so s50 means min_score=0.5; fcs is 0–100 so s50 means min_score=50.

Format

JSONL shards named *.jsonl.shardKofN (not .jsonl, so load_dataset will not auto-parse them). One sample per line:

{
  "messages": [ {"role": "system", "...": "..."}, ...,
                {"role": "assistant", "...": "THE TRAINING TARGET"} ],
  "n_ctx": 12,
  "n_tokens": 21254,
  "task_id": "arc_agi_2_00576224",
  "source": "student",
  "iter": "iter_0001",
  "ctx_trimmed": false
}

Token ids are not stored. Render with the Qwen3.5 chat template at load time and pass tools.json (four functions: Bash / Read / Write / Edit) — it is part of the context the agent saw.

from pathlib import Path
import json
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

root = Path(snapshot_download("REPO_ID", repo_type="dataset"))
tools = json.loads((root / "tools.json").read_text())
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B")

def rows(corpus: str):
    for p in sorted((root / corpus).glob("*.jsonl.shard*")):
        if p.name.endswith(".manifest.json"):
            continue
        with p.open() as f:
            for line in f:
                yield json.loads(line)

sample = next(rows("arc2_48k_best-s0_64k"))
ctx  = tok.apply_chat_template(sample["messages"][:sample["n_ctx"]], tools=tools, tokenize=False)
full = tok.apply_chat_template(sample["messages"], tools=tools, tokenize=False)

Every shard has a sibling .manifest.json with counts, rejects, and sha256 pins of AGENT_SYSTEM.md / ralph_agent.py.

Two things to know before training

1. Reasoning is a target, never context. Prior-turn reasoning_content was dropped from history at inference. Only the final target turn in a sample carries it. Reintroducing think blocks into context inflates length ~30% and diverges from inference.

2. Teacher (GLM-5.2-Int4) reasoning is sparse. Empty reasoning_content renders as a degenerate <think></think> and trains "act without thinking". Do not mix student and teacher naively.

Licence

ARC-AGI-2 task content: arcprize/ARC-AGI-2. frontier-cs-algo: the FCS release licence. Model outputs: Qwen3.5 and GLM-5.2 terms. Grids and held-out test answers are not included.