yibaiershi_cont — reformatted Affine SN120 teacher SFT set

63,857 prompt–completion pairs over 23,478 distinct agent turns and 1,712 repositories, distilled from the public duel records of Affine (Bittensor SN120).

This is a reformat of iamPi/yibaiershi_cont, not a new harvest. All credit for collecting and filtering the underlying rollouts belongs there. The changes here are the schema and the answer format — see What changed below.

Fields

Denormalized: one self-contained row per sample, no join required.

field	type	description
prefix	list[{role, content}]	chat history — the prompt
turn_id	string	"{traj_id}:{turn_idx}"; split with rpartition(":")
z_raw	string	teacher reasoning, exactly as sampled
y_raw	string	teacher action — a closed ```bash block
answer_raw	string	assistant body — the completion
messages	list[{role, content}]	prefix + the assistant reply, one transcript

Answer format

answer_raw (and the final messages turn) is plain text with no <think> channel:

THOUGHT: {reasoning}\n\n```bash\n{command}\n```

Verified on all 63,857 rows: starts with THOUGHT: , contains no think tags, messages[-1]["content"] == answer_raw, messages[:-1] == prefix, and the action parses as a closed bash block.

from datasets import load_dataset
ds = load_dataset("vuhaian/yibaiershi_cont", split="train")
# hand ds["messages"] to SFTTrainer and let the chat template render it

Pass add_special_tokens=False when tokenizing if your template already emits BOS-like markers (GLM emits [gMASK]<sop>), and set max_length high — prompts run to 32k tokens, so a low default silently drops most examples.

What changed from the upstream dataset

1. Denormalized. Upstream ships sft_rows.jsonl + sft_turns.jsonl to be joined on turn_id; here the prefix is inlined per row. Costs disk, removes a step.
2. Think tags dropped from the completion. Upstream stores the subnet's scored body, </think>\nTHOUGHT: {z}\n\n{y} — it opens with a closing tag because the validator's gen_prompt() ends the prompt inside an open <think>. That form is only correct if you concatenate it onto a <think>-terminated prompt; rendered through a chat template it produces a nested or dangling tag. This release stores the model-agnostic plain form. Stray tags the teacher emitted mid-generation were also removed (~23k rows in z, 3 inside a bash command).
3. messages column added, so the set is usable directly as a conversational SFT dataset.

z_raw and y_raw are carried through untouched, so the exact body the validator scored is always rebuildable:

z = row["z_raw"].strip()
if z.startswith("<think>"):
    z = z[len("<think>"):]
z = z.replace("</think>", "").strip()
scored_body = "</think>\nTHOUGHT: " + z + "\n\n" + row["y_raw"]

Composition

• 63,857 rows over 23,478 distinct turns (mean 2.7 rows/turn)
• 1,712 repositories; conan-io/conan 13.5%, pygments/pygments 4.3%, then a long tail — no other repo above 0.7%
• corpus epoch 10, manifest 45900875e046; every turn_id resolves in the public corpus

Notes before training

• Duplicate prompts are intentional. The teacher samples at temperature 0.8, so one turn yields several distinct actions. The multiplicity is the teacher's action distribution. Deduplicate only if you specifically want a one-sample-per-turn set.
• Group by turn_id before splitting train/eval — 63,857 rows cover 23,478 turns, so a random row split puts the same prefix on both sides.
• The repo mix is skewed by construction. Duel turn selection round-robins over repo|phase strata and the phase key splits on PR number, so a repo's weight tracks its PR count, not its data volume. This mirrors the distribution models are actually scored against, which may or may not be what you want.
• Upstream's filters are inherited: thoughts under 80 chars, non-positive teacher self-lift, exact duplicates and unparsable actions were already dropped there.