# Data and Environment Curation for Post-Training LLMs — Mahesh Sathiamoorthy, Bespoke Labs

Source: [AI Engineer](https://www.youtube.com/watch?v=ewtOo0scUh0)  
Feed7 permalink: https://feed7.dev/p/data-and-environment-curation-for-post-training-llms-mahesh-sathiamoorth-19ab77a  
Published: 2026-07-31T22:00:06.000Z  
Trust: Source Linked (source_linked)

## Why Included

Post-training gains depend heavily on task selection, rollout quality, and environment design. For many enterprise agents, curated SFT may deliver most of the value before costly RL.

## Source Summary

Bespoke’s curation recipe selects source prompts, mixes and filters them, generates teacher answers, then filters again. Across its work, **multiple answers per question** helped, while the strongest model was not always the best teacher.

## Practical Implication

Treat data recipes and environments as versioned engineering assets. Start with **SFT** for the required behavior, run ablations at each curation stage, and reserve **RL** for gains that justify its added compute and infrastructure.

## Agent-Ready Context

Bespoke’s curation recipe selects source prompts, mixes and filters them, generates teacher answers, then filters again. Across its work, **multiple answers per question** helped, while the strongest model was not always the best teacher.

Treat data recipes and environments as versioned engineering assets. Start with **SFT** for the required behavior, run ablations at each curation stage, and reserve **RL** for gains that justify its added compute and infrastructure.

Synthetic rewriting and task augmentation did not reliably help in the reported agent work. Production datasets can also be imbalanced, so fine-tuning may amplify rare-looking attributes unless the mix and outputs are checked carefully.

## Connected Context

Feed7 judgment across 318 accumulated Signals:

This turns post-training data and environments into versioned, ablated engineering systems rather than assuming more synthetic data, a stronger teacher, or RL will improve agents. Against the candidates’ emphasis on runtime harnesses and controls, it adds an upstream reliability constraint: prompt mix, answer diversity, filtering, and production imbalance can determine whether those systems receive useful behavior at all.

- [OpenForgeRL: Train Harness-native Agents in Any Environment](https://feed7.dev/p/2607-21557v1-0blvz16) — OpenForgeRL supplies deployment-harness-native RL infrastructure, while this signal narrows when that machinery is warranted by recommending curated SFT first and RL only for gains that justify its extra cost.
- [Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning](https://feed7.dev/p/scaling-to-long-horizons-ross-taylor-chengxi-taylor-general-reasoning-0jwtg4d) — The long-horizon work details RL’s credit, context, and scheduling costs; this signal reinforces treating those costs as a reason to establish SFT gains and stage-by-stage ablations before adopting RL.
- [Why Off-the-Shelf AI Doesn't Understand Money — Udi Menkes, Intuit](https://feed7.dev/p/why-off-the-shelf-ai-doesn-t-understand-money-udi-menkes-intuit-0y6w9rk) — Intuit requires valid state-action-outcome evidence for domain recommendations; this signal broadens that data-quality prerequisite to the full training recipe, including source mix, teacher outputs, filtering, and imbalance checks.

## Context Map

- Layer: agent
- Domains: coding, data
- Topics: harness-engineering, agent-reliability, sandboxing

## Uncertainty

- Synthetic rewriting and task augmentation did not reliably help in the reported agent work. Production datasets can also be imbalanced, so fine-tuning may amplify rare-looking attributes unless the mix and outputs are checked carefully.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.