QandyAI/qandy-1.0-27B-data

Qandy 1.0 27B: Training and Evaluation Data

The data used to train and evaluate Qandy 1.0 27B, from Starburst, QANDY's trading research line. Each example pairs a compact market snapshot for one crypto perpetual-futures symbol with a decision: LONG, SHORT or NO TRADE. Trades come with leverage, a stop, a target and key levels, plus a step-by-step reasoning trace that cites only the signals in the snapshot.

The examples come from a live, high-leverage scalping desk. Losing trades and costly skips are kept and labelled rather than filtered out, so the data can teach a model what not to do as well as what to do.

Config	Split	Rows	Used for
sft	train	896	Stage 1 supervised fine-tuning
kto	train	1,384	Stage 2 KTO preference tuning
holdout_real	test	182	Main evaluation on live desk outcomes
backfill_10k	test	10,000	Larger replay evaluation

from datasets import load_dataset

sft = load_dataset("QandyAI/qandy-1.0-27B-data", "sft", split="train")
kto = load_dataset("QandyAI/qandy-1.0-27B-data", "kto", split="train")
holdout = load_dataset("QandyAI/qandy-1.0-27B-data", "holdout_real", split="test")

Snapshot format

Every prompt is a two-message chat: a fixed system prompt and a user snapshot.

You are qwopus-trades, the entry-reasoning model for a crypto perp desk. From the snapshot, decide LONG, SHORT, or NO TRADE — reason through the signals, then give one decision with leverage, a risk plan (stop, target), and key levels (support below price, resistance above price). Cite only signals in the snapshot.

The snapshot carries the symbol, a reference price, and whichever of these signals were available at decision time: RSI (5m/15m/1h/4h), MACD 4h, Ichimoku, VWAP, funding versus its 70th percentile, open interest, spot and perp CVD, taker flow, whale flow, retail and smart-money long/short ratios, liquidation distance, support and resistance, and a one-line context. Fields that were missing at decision time are left out or shown as -.

A few rows also end with a desk constraint, for example [ACCOUNT] Free margin 15.2% (min 30% to open a new position). In those rows the correct answer is NO TRADE whatever the signals say.

The assistant turn is a <think> trace followed by one decision line:

Decision: SHORT ETH/USDT LIMIT @ 2929.58 at 30x. Setup: breakdown. Conviction: medium. Risk plan: stop 2947, target 2868.77. Key levels: Support 2916.15 | Resistance 2988.23.
Decision: NO TRADE. Conviction: low. Key levels: Support 1680.03 | Resistance 1702.88.

Configs

sft

{"messages": [system, user, assistant]}. There are 896 examples: 693 winning trades (417 SHORT, 276 LONG) and 203 good skips (NO TRADE). Skips were sampled at 30% of the winner count. Symbols are BTC/USDT 359, ETH/USDT 358, SOL/USDT 113 and AVAX/USDT 66.

kto

{"prompt": [system, user], "completion": [assistant], "label": bool}, the unpaired format used by TRL's KTOTrainer.

• Desirable (true), 896 rows: the same examples as sft.
• Undesirable (false), 488 rows:
  • 458 real losing trades (280 SHORT, 178 LONG), held out of SFT
  • 30 missed winners: skips whose counterfactual trade would have reached at least +1R. These are relabelled as undesirable so that skipping a winning setup is penalised.

Counterfactual method: the trade direction is a vote of MACD 4h, RSI 4h and RSI 1h. The stop is the cited support or resistance and the target is ±1.5R, walked forward over a 6-hour horizon and checked against Binance 1-minute prices. If the stop and the target are both hit inside the same candle, it counts as stopped out. The vote under-counts oversold-bounce longs, so 30 is a lower bound.

Qandy 1.0 27B used undesirable_weight 1.53 to balance the two classes.

holdout_real

182 live desk setups that the model never trained on:

• record_type: resolved (83: 52 wins, 31 losses) or skip (99, where the desk stood aside)
• output.parsed.decision: the desk leader's decision
• outcome: for resolved rows, status (win or loss), signed r_multiple and an ambiguous flag. Empty for skips.

50 rows carry an older wording of the system prompt, exactly as it was logged at the time. They were kept unchanged so that results stay comparable with the reported evaluation.

backfill_10k

10,000 replayed desk cycles on BTCUSDT and ETHUSDT (5,000 each), from 2025-12-10 to 2026-06-07. The leader decisions were produced by claude-opus-4-8.

• record_type: resolved 1,720, unfilled 2,360 (limit order never filled), skip 5,920
• output: the leader's decision_line and reasoning
• outcome: fill and P&L details for resolved rows (entry, fill_time, leverage, risk_pct, reward_pct, r_multiple, pnl_price_pct, pnl_margin_pct, fee_margin_pct), and the intended entry for unfilled rows
• Metadata: idx, source, symbol, cycle_time, ref_price, kto_label

These snapshots use an older replay format, so they include Ichimoku and retail long/short fields but no whale flow, and are not identical to the live prompt shape. Use holdout_real for production-faithful evaluation.

Evaluation protocol

Each model's decision on a setup is compared with the leader's trade:

• Win retention: share of the leader's winners where the model took the same side
• Loss avoidance: share of the leader's losers the model skipped or flipped
• Gated sum R: the leader's P&L, in R multiples, if only the trades the model approved were taken

The scoring script is scripts/eval_holdout.py in the model repo. That repo's card has the results.

Leakage check

No user snapshot in holdout_real or backfill_10k appears verbatim in sft or kto.

Limitations

• Small and narrow. The data covers about six months (December 2025 to June 2026) of one desk's style: high-leverage scalping on four perpetual pairs, mostly BTC and ETH. It isn't a general market dataset.
• Desk-specific labels. "Good" and "bad" mean good or bad for this desk's rules and risk settings, not universally.
• Sparse signals. Many snapshots lack several fields, and reasoning traces sometimes lean on the few that are present.
• Small evaluation set. holdout_real has 83 resolved trades, so metrics on it are noisy (roughly ±2–3R).
• Not financial advice. This data is for research. Don't use models trained on it to place real orders without independent risk controls.

About QANDY

QANDY builds personal financial agents. Learn more at qandy.ai.