Castform RL fine-tune — what we stole

Source: castform.com — RL fine-tune SaaS that lets engineers train open-source models on proprietary data + a user-defined reward function. Visited 2026-06-13; surface inspected via product copy + the public Python SDK sketches.

Why look at them: Castform’s product thesis (“4B fine-tuned beats GPT-5.4 on the narrow task at 1.0× the cost”) is identical to TinyGPT’s Mac-specialist thesis. The interesting question isn’t whether small models can win — it’s which engineering primitives make training them practical. Castform exposes four that are worth borrowing.

This page is a steal-map (same format as agent-context-hierarchy.md): one section per pattern, what it gives us, where it goes in the PRDs.

The four steals

1. Composite reward functions — filed as B28

Castform: a reward is a composition of named dimensions — correctness + conciseness + citation + tool_call_efficiency — each independently scorable, then aggregated via weights into the training signal. The dashboard shows the per-dimension breakdown per rollout so you can see which axis is driving the gradient.

Where we are: DPO has a single implicit reward from log_π_pol − log_π_ref. ES has a single scalar (negative loss). There’s no abstraction for “this reward is actually four things in a trenchcoat”, which is the practical shape every real specialist eventually needs.

Filed as B28 composite-reward-framework in PLAN.md — a Reward struct with named dimensions, weighted aggregation, per-dimension logging. Usable from DPO (as the chosen/rejected score), GRPO (5.1), and ES.

The base struct ships in this PR (CompositeReward.swift). The training-loop integrations are the rest of B28.

2. Trace-driven dataset synthesis — filed as B29

Castform: pull from production agent traces (Braintrust, Langfuse, LangSmith) and RAG corpora (Turbopuffer, Pinecone, Chroma, Postgres). Auto-filter via dedup + tool-echo drop + LLM-judge pivot at configurable thresholds (0.6–0.9). The output is a training-ready JSONL the user never had to hand-label.

Where we are: B22 (token-preserving trajectory recorder) ships the substrate — every .atraj file carries input_ids, output_ids, tool calls, rewards. But there’s no consumer that turns the substrate into SFT/DPO training data.

Filed as B29 trace-to-training-data in PLAN.md — the bridge between B22 and A1. Reads .atraj files, runs the existing tinygpt dedupe, tinygpt judge (E7) shims for filtering, emits training JSONL.

3. Multi-hop reasoning classification — filed as B30

Castform: classify training prompts by reasoning depth (single-hop, multi-hop, comparison) so the training mix is balanced — too much single-hop and the model never learns to chain. Too much multi-hop on a base that hasn’t shipped basic capability yet and it diverges.

Where we are: no classifier; balance is per-corpus by hand. Works for narrow specialists; falls over for the agent-trace data B29 will produce, which is intrinsically mixed-depth.

Filed as B30 prompt-reasoning-classifier — small classifier head trained on a labeled subset, scores any prompt into {single-hop, multi-hop, comparison, other}. Drops into B29’s filtering pipeline + the leaderboard’s per-category breakdown.

4. Pluggable BaseEnv interface — not filed (out of scope)

Castform: users subclass BaseEnv with async run_tool() + compute_reward(). The platform handles rollouts + the training loop. This is a UX choice for a SaaS surface.

Where we are: TinyGPT is a CLI + Swift library, not a SaaS. Users write a Swift EvalCompare.Row row or a Python eval harness; the “plug in your env” abstraction is already there, just spelled in two languages.

Deliberately not filed. The B6 Mac app (Factory tab) is the right place for a friendlier env-config UX, and B6 already covers that — adding a BaseEnv Python class layer is fake reuse since the underlying mechanism is the same.

What we deliberately did not steal

• The dashboard. Castform’s training dashboard (avg reward per step, response length, solve rate, max reward per prompt) is covered by C10 train-run-dashboard + B23 agent-eval-protocol — already filed. Castform’s specific axes are listed in B28’s acceptance criteria so the train viewer adds those columns when composite reward is on.
• The pay-per-compute UX. Castform is a SaaS; we’re a Mac CLI. Not transferable.
• External observability tool ingestion (Braintrust / Langfuse / LangSmith). B29 reads our own .atraj files. External-tool ingest is a V2 if users actually have agent traces in those systems they want to import; until then, premature.

The addition this analysis surfaced (our note)

Castform’s “solve rate (pass@k)” metric is the per-prompt version of what B23 protocol does across the eval suite. Combined: every leaderboard row reports mean ± σ (B23) AND the per-prompt pass@k distribution under the recorded budget. That’s a deliverable for B23’s V2 — not blocking now, but the data is free once K-pass rollouts are running.