data-and-rewards
Validate VLM-R1 JSONL datasets, map image roots, and choose or test reward methods.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-and-rewards
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Validate VLM-R1 JSONL datasets, map image roots, and choose or test reward methods.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/data-formats.md3.7 KB
- 参考文档references/reward-functions.md3.2 KB
- 参考文档references/troubleshooting.md2.6 KB
- 脚本scripts/score_bbox_outputs.py10.8 KB
- 脚本scripts/validate_jsonl_dataset.py12.5 KB
03 SKILL.md 正文
data-and-rewards
Use this sub-skill when you need to:
- validate single-image, multi-image, or text-only JSONL rows
- map relative image paths against paired image roots
- choose between file-level
reward_methodand row-levelaccu_reward_method - explain answer formats for label, math, bbox, and OD fixtures
- run lightweight bbox scoring on tiny synthetic examples
Start here
- Read
references/data-formats.md. - Read
references/reward-functions.md. - Read
references/troubleshooting.md. - Run
scripts/validate_jsonl_dataset.pyfor schema checks. - Run
scripts/score_bbox_outputs.pyfor tiny bbox fixtures.
Route elsewhere
- Training command construction, launch flags, or DeepSpeed details ->
training-workflows - Model module interfaces or processor wiring ->
model-modules - Saved-output evaluation workflows ->
evaluation
What this sub-skill covers
- one JSON object per line
imageas a string or list of stringsconversationswith a question turn and an answer turn- colon-separated file, image-root, and method pairing
- reward registry names:
accuracy,format,length,repetition - accuracy methods:
default,mcq,yes_no,llm,map,math,weighted_sum,od_ap,od_ap50,odLength,all_match
Guardrails
- Keep image paths relative to the paired image root.
- Keep answers concise; the loader wraps the stored answer in
<answer>...</answer>internally. - Do not tell future agents to inspect the source repository for these rules.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).