> Canonical page: https://posttrainllm.com/studies Tracked external evidence # What we studied, what we kept, and what we rejected. Every external product, model, platform, writeup, and case study that changed the PostTrainLLM learning map gets one canonical record. Each dossier keeps the local translation beside its evidence limits so a useful idea never turns into an unsupported claim. **Canonical studies** 20 **Distinct dispositions** 11 **Unresolved ownership** 0 Evidence index ## All retained studies Indexed from the tracked review ledger and industry learning map. product review partially adopted ### [TrainLoop AI](https://posttrainllm.com/studies/trainloop-ai) Failed-attempt accounting, slice metrics, trace review, candidate selection, batch-first post-training, policy lag, and LoRA update geometry make autoresearch inspectable instead of hiding the search process behind a final score. [Open evidence dossier →](https://posttrainllm.com/studies/trainloop-ai) case-study review scaffolded ### [TrainLoop NollaMD case study](https://posttrainllm.com/studies/trainloop-nollamd) Candidate-selection framing can make a sparse task measurable when unconstrained generation is too difficult or too ambiguous to grade reliably. [Open evidence dossier →](https://posttrainllm.com/studies/trainloop-nollamd) technical writeup scaffolded ### [TrainLoop OAPL writeup](https://posttrainllm.com/studies/trainloop-oapl) Batch rollouts, offline scoring, and compact updates can make a post-training loop easier to inspect, while policy lag can stabilize learning under some reward surfaces. [Open evidence dossier →](https://posttrainllm.com/studies/trainloop-oapl) technical writeup scaffolded ### [TrainLoop LoRA geometry writeup](https://posttrainllm.com/studies/trainloop-lora-geometry) Effective update geometry, layer placement, and controlled rank analysis can explain why two adapters with similar parameter counts behave differently. [Open evidence dossier →](https://posttrainllm.com/studies/trainloop-lora-geometry) product writeup parked ### [TrainLoop Mercor coding-agent post](https://posttrainllm.com/studies/trainloop-mercor-coding-agent) File-native and code-native harnesses may fit knowledge-work agents better than bespoke tool APIs because the environment already exposes durable state and familiar operations. [Open evidence dossier →](https://posttrainllm.com/studies/trainloop-mercor-coding-agent) product positioning adopted ### [Baseten post-training positioning](https://posttrainllm.com/studies/baseten-post-training) Post-training is a system of custom data, reward shaping, performance work, infrastructure, evaluation, and packaging rather than a generic fine-tuning screen. [Open evidence dossier →](https://posttrainllm.com/studies/baseten-post-training) model landscape tried ### [Hugging Face model hub and SQL specialists](https://posttrainllm.com/studies/hugging-face-sql-specialists) Small public SQL models provide realistic baselines and expose how much benchmark, model size, and serving format affect apparently similar claims. [Open evidence dossier →](https://posttrainllm.com/studies/hugging-face-sql-specialists) model landscape adopted as evaluation guidance ### [Defog SQLCoder and Arctic Text2SQL class models](https://posttrainllm.com/studies/defog-sqlcoder-arctic-text2sql) Strong SQL systems report execution accuracy on substantial SQL benchmarks rather than relying on exact string match alone. [Open evidence dossier →](https://posttrainllm.com/studies/defog-sqlcoder-arctic-text2sql) platform review rejected for core capability ### [Apple on-device Foundation Models](https://posttrainllm.com/studies/apple-on-device-foundation-models) Apple's system model is a free and private routing floor, but its context and action-grounding limits make it unsuitable as the core capability dependency for a realistic tool catalog. [Open evidence dossier →](https://posttrainllm.com/studies/apple-on-device-foundation-models) product review partially adopted ### [Castform RL fine-tune platform](https://posttrainllm.com/studies/castform-rl-finetune) Composite rewards, trace-driven data loops, reasoning-depth classification, and explicit environment contracts make reinforcement-style fine-tuning more inspectable. [Open evidence dossier →](https://posttrainllm.com/studies/castform-rl-finetune) agent architecture review partially adopted ### [Cline and agent context hierarchy research](https://posttrainllm.com/studies/cline-context-hierarchy) Structured-output enforcement and an explicit context hierarchy can reduce ambiguous agent actions and make tool boundaries easier to inspect. [Open evidence dossier →](https://posttrainllm.com/studies/cline-context-hierarchy) tokenization system review parked ### [Gigatoken](https://posttrainllm.com/studies/gigatoken) SIMD and cache-heavy bulk BPE can make first-pass tokenization much faster on large corpora, including Qwen tokenizers on Apple Silicon. [Open evidence dossier →](https://posttrainllm.com/studies/gigatoken) model review and local smoke reviewed and rejected ### [Needle 2](https://posttrainllm.com/studies/needle-2) A compact call-only model combines top-five tool retrieval, constrained decoding, confidence escalation, bounded context, quantization, and a small packaged Mac artifact. [Open evidence dossier →](https://posttrainllm.com/studies/needle-2) runtime review and local smoke validated proof ### [parakeet.wgsl](https://posttrainllm.com/studies/parakeet-wgsl) Raw WebGPU plus SIMD-WASM can run Parakeet TDT browser inference with public kernels, a package format, cache behavior, a converter, and a benchmark harness. [Open evidence dossier →](https://posttrainllm.com/studies/parakeet-wgsl) industry case study study only ### [Savante and Aryabhata](https://posttrainllm.com/studies/savante-aryabhata) Curated domain data, filtered reasoning traces, supervised fine-tuning, model merging, and verifiable-reward reinforcement learning combine in a JEE mathematics specialist. [Open evidence dossier →](https://posttrainllm.com/studies/savante-aryabhata) industry case study study only ### [Bonsai 2 27B](https://posttrainllm.com/studies/bonsai-2-27b) Ternary representation, packing overhead, activation transforms, kernel support, artifact size, runtime memory, and capability retention can rank a model differently depending on the deployment constraint. [Open evidence dossier →](https://posttrainllm.com/studies/bonsai-2-27b) industry case study learning queued ### [QORL](https://posttrainllm.com/studies/qorl) Parameter-aware query optimization needs equal search budgets, fresh held-out measurement, SQL-equivalence checks, and a native-planner baseline before an offline plan search can claim value. [Open evidence dossier →](https://posttrainllm.com/studies/qorl) industry case study study only ### [Inside vLLM](https://posttrainllm.com/studies/inside-vllm) Admission control, scheduling, paged KV allocation, prefill, decode, continuous batching, caching, and distributed execution explain why a serving engine behaves differently from a single-request model loop. [Open evidence dossier →](https://posttrainllm.com/studies/inside-vllm) industry case study study only ### [Splash model-specific Mac inference](https://posttrainllm.com/studies/splash-mac-inference) Shape-specific Metal kernels, a model-specific speculative draft, packed weights, cache reuse, and a startup memory plan trade generic model support for Mac-local serving performance. [Open evidence dossier →](https://posttrainllm.com/studies/splash-mac-inference) training-framework review partially adopted ### [Halo / White Circle](https://posttrainllm.com/studies/halo) HF-native training and portable checkpoints, independent expert/context/tensor/data parallelism, faithful multi-turn trajectories, operation-specific model support, and explicit resume/export/serving boundaries make the single-machine-to-cluster transition inspectable. [Open evidence dossier →](https://posttrainllm.com/studies/halo)