forge-neoforge-sft-v2.0.2 (full)

Supervised fine-tuning corpus for Forge 1.20.1 -> NeoForge 1.21.1 mod migration. Every task is grounded in exact pinned upstream snapshots (branch/tag + commit sha per record): the model is never asked about invisible code - all context files are serialized inline (inline-fenced-v2) and their presence is enforced by a blocking gate (G-CTX).

7030 records / 15 families / 631 repositories. Carries the entire v2.0.1 release (3615 records) byte-identical, expanded with: (1) discovery re-scan - full ls-remote ref enumeration of 757 previously unmatched repositories, converting 141 of them into valid pin pairs (plus 43 registry leftovers; 184 new pairs total), and (2) a PR-thread multiturn layer - repair dialogues mined from real merged PRs via per-commit .patch diffs and exact git-object file states. multiturn_repair grows 9 -> 90.

Splits (repo-disjoint)

split	records	repos
train	6278	555
validation	419	37
eval	333	39

Repositories are disjoint across splits (no leakage). eval is slightly above the 300 target because split purity is preserved per repository.

Families

family	total	train / val / eval	description
api_version_lookup	284	265 / 13 / 6	symbol-level Forge->NeoForge replacement lookups validated literally in the repository own pre/post snapshots
code_generation	340	299 / 19 / 22	implement a new class from scratch consistent with attached conventions; the answer is the real upstream-added file
config_and_resources	843	741 / 53 / 49	lang/pack.mcmeta/toml/recipe/tag updates between pins, real diff notes only
context_verification	19	14 / 2 / 3	context adequacy checks (v2 family, extended)
grounded_qa	793	700 / 51 / 42	code comprehension Q&A grounded in pinned files (events, registrations, config, responsibilities)
insufficient_context	460	408 / 27 / 25	symmetric negative cases: missing internal dependencies, version-mismatch claims, unconfirmed third-party APIs - the correct behavior is to refuse and request exactly what is needed
issue_bugfix	205	178 / 6 / 21	real upstream bugfixes: prompt carries the exact pre-fix parent revision; the answer is the real fix commit result
issue_context_check	653	599 / 27 / 27	real issue triage: version-conflict reports and insufficient bug reports (no logs) -> request specifics, do not act
multiturn_repair	90	79 / 7 / 4	real repair dialogues: PR-thread sequences (merged PRs, per-commit diffs, optional verbatim reviewer quotes) and fix -> follow-up commit chains; 4-message
porting	1647	1471 / 113 / 63	L1 file migration: migrate an exact Forge 1.20.1 snapshot of one file to NeoForge 1.21.1; the answer is the upstream 1.21.1 state of that file (pinned sha256)
porting_diff_review	772	697 / 49 / 26	L1: review a proposed single-file migration diff that IS the real upstream change; verify hunks and state what a full port additionally requires
pr_review	374	340 / 16 / 18	file-by-file review of real merged PR diffs with verifiable observations
project_porting	199	177 / 14 / 8	L2: subsystem-level migration; multiple dependent files planned, migrated and cross-referenced
review	262	225 / 19 / 18	L1 review-and-verify: analyze real 1.20.1 code / real migration artifacts, grounded quotes only
test_generation	89	85 / 3 / 1	game-test skeletons grounded in real registrations + NeoForge 1.21.1 gametest API surface; honest static_checked status

Methodology (honesty core)

• Pins. Each record carries pins (branch or tag + commit sha for Forge 1.20.1 and NeoForge 1.21.1) and upstream_truth (paths + sha256 of the exact snapshots it was derived from).
• Discovery re-scan (new). The original registry marked 753 repos missing_side using fixed branch-name patterns. Full git ls-remote ref enumeration (branches + tags + HEAD) with flexible MC-version matching produced 495 candidate repos; each pair was then verified by the same full-tree loader gate plus a new MC-version content check (1.20.1 / 1.21.1 must literally appear in build/asset files at the respective pin). 61 candidates were honestly rejected (35 fabric targets, 22 missing version text, 4 forge-less), 2 had no java diff.
• PR-thread multiturn (new). For merged PRs, https://github.com/.../pull/N.patch (non-API endpoint, no quota) provides the real per-commit diffs and commit messages. The dialogue uses commit 1 as the initial change and the last java-touching commit as the follow-up correction; file states are exact git objects at the PR commit shas (shallow fetch, depth=2). Where the core API allowed, real reviewer comments are quoted verbatim with author + URL; on a shared IP the quota was exhausted, so most follow-up turns quote the real follow-up commit message instead - labeled as such, never paraphrased into fake reviewer speech.
• No fabricated builds. Statuses: static_checked (static verification against pinned snapshots). validation.method states exactly what was verified. No compile/run was performed in the producing environment for this release.
• Gates. G1 unique ids (duplicate-id records dropped); G-CTX (every context file byte-present in a user turn - blocking); G3 context sha256; G5 secret scan; G9 carried v2.0.1 answers byte-identical (3615/3615) and every carried v2.0.1 file an exact byte-prefix of the merged file; G10 upstream re-fetch sample (12 sha256 matches, 2 skipped by design); exact-duplicate + duplicate-id drop (3 records); repo-level split disjointness.
• Loader gate. Full-tree git grep at BOTH pins for all 247 candidate repos this round: 184 pass (129 multiloader with fabric/quilt subtrees excluded), 61 rejected, 2 without java diff.
• Multiloader policy (v2 precedent): repos with Common/Forge/NeoForge trees are valid material; fabric/quilt subtrees are excluded from every selection.

Record schema (per record)

task_id, task_family, task_type, source_target, prompt, context_files[] (path, content, sha256, source_url), messages (self-contained inline-fenced-v2; for multiturn_repair a 4-message dialogue), expected_changes, evidence[] (claim, source_url, source_revision), build_environment, validation (honest status), repo, pins, split, serialization, upstream_truth (path/sha256/revision; PR dialogues carry pr_number + commit shas).

Limitations

• compile_verified statuses are absent: NeoForm decompile requires >=8 GB RAM (Vineflower -Xmx4G), unavailable in the producing environment; nothing was fabricated to hide this.
• Reviewer-quote multiturn subset is thin: GitHub core API (60/h, shared IP) was exhausted; PR-thread follow-ups are quoted from real follow-up commit messages instead.
• test_generation records are grounded skeletons (API surface verified against NeoForge 1.21.1 tree), not compiled tests.
• Some discovered pairs pin release tags rather than branch tips (provenance keeps the exact ref + sha); tags were required when a repo had no dedicated 1.21.1 branch.
• Issue/PR texts are used as real user reports; issue-layer records are limited to repos with fetchable commit objects.

Source & provenance

• Base: ComerX/forge-neoforge-sft-v2.0.1 (3615 records, carried byte-identical) <- ComerX/forge-neoforge-sft-v2 (1070 records).
• Expansion round: 184 new repository pairs (43 unused registry + 141 discovered by ls-remote re-scan of 757 missing_side repos); issues/PRs via GitHub search for the new repos.
• Upstream licenses: per-repository (MIT/Apache-2.0/LGPL-2.1/BSD and similar); pins allow byte-exact verification of every derived answer.

Citation

@dataset{comerx_sft_v202_2026,
  title  = {forge-neoforge-sft-v2.0.2 (full): Forge 1.20.1 -> NeoForge 1.21.1 SFT corpus},
  author = {ComerX},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ComerX/forge-neoforge-sft-v2.0.2}
}