Buckets:

qgallouedec
/

ecosystem-watch

Files

xet

qgallouedec/ecosystem-watch / sources.json

qgallouedec

3 months ago

download

raw

11.3 kB

{
  "_comment": [
    "Configuration for the TRL ecosystem intelligence bot.",
    "The bot runs as a Hugging Face scheduled Job. Everything it needs lives in the bucket",
    "hf://buckets/qgallouedec/ecosystem-watch, mounted read-write at /eco:",
    "    sources.json  this file    prompt.txt  the bot's instructions",
    "    run.sh        the Job entrypoint",
    "    state/        seen.jsonl dedupe ledger + watchlist.jsonl + last_run stamp",
    "    logs/         one transcript per run, named by UTC start time",
    "Edit this file to add/remove sources — the prompt reads it at runtime, nothing else needs changing.",
    "tier 1 = deep scan (releases, merged PRs, open PRs, issues, discussions, milestones).",
    "tier 2 = shallow scan (releases + notable/labelled activity only).",
    "'focus' tells the bot what kind of change in that repo actually matters to TRL."
  ],

  "_slack_setup": [
    "This bot posts to its OWN channel, separate from the issue-triage bot.",
    "One-time setup, done by a human (the bot token has chat:write but not channels:manage,",
    "so it cannot create a channel or look one up by name):",
    "  1. In Slack, create the channel (suggested: #trl-ecosystem).",
    "  2. In that channel, run: /invite @trl_triage_bot",
    "     Without this, chat.postMessage fails with not_in_channel.",
    "  3. Copy the channel ID (channel name > View channel details > bottom of the About tab,",
    "     or the last path segment of the channel URL — it starts with C).",
    "  4. Paste it into slack_channel below, replacing REPLACE_ME.",
    "",
    "run.sh preflights this and exits without burning an API run if it is unset."
  ],
  "slack_channel": "C0BLEAM73KK",
  "lookback_hours_default": 30,
  "max_items": 6,

  "github": [
    { "repo": "huggingface/transformers", "tier": 1,
      "focus": "Trainer/TrainingArguments API changes, deprecations, model loading (dtype/device_map/quantization), attention implementations & masking, cache classes, chat templates & processors, generation config, tokenizer behaviour, anything that breaks TRL imports or silently changes loss/logits." },
    { "repo": "huggingface/accelerate", "tier": 1,
      "focus": "FSDP2/DeepSpeed integration, gradient accumulation semantics, dataloader sharding, context parallelism, plugin/config schema changes, prepare() behaviour." },
    { "repo": "vllm-project/vllm", "tier": 1,
      "focus": "Releases and breaking API changes in LLM/AsyncLLM, sleep/wake_up & weight-hot-swap paths, collective_rpc, V1 engine internals TRL depends on for the vLLM backend, logprobs semantics, server/OpenAI-API changes." },
    { "repo": "huggingface/peft", "tier": 1,
      "focus": "LoRA/adapter API changes, merge/unmerge and adapter hot-swap, new PEFT methods worth exposing in TRL configs." },
    { "repo": "huggingface/datasets", "tier": 1,
      "focus": "Dataset/IterableDataset API changes, map/column semantics, streaming, anything affecting TRL dataset formats or collators." },
    { "repo": "huggingface/huggingface_hub", "tier": 1,
      "focus": "Hub upload/push_to_hub, model cards, inference/jobs API, auth changes used by TRL." },

    { "repo": "deepspeedai/DeepSpeed", "tier": 2,
      "focus": "ZeRO-3 gather/partition behaviour, memory regressions, compatibility with recent transformers/accelerate." },
    { "repo": "pytorch/pytorch", "tier": 2,
      "focus": "Release notes and deprecations touching torch.compile, FSDP2, DTensor, autocast/dtype APIs, distributed primitives." },
    { "repo": "pytorch/ao", "tier": 2,
      "focus": "Quantized/low-precision training (QAT, fp8, NVFP4) usable from TRL." },
    { "repo": "linkedin/Liger-Kernel", "tier": 2,
      "focus": "Fused loss kernels (chunked CE, DPO/GRPO losses) TRL wires up, and their API/signature changes." },
    { "repo": "huggingface/kernels", "tier": 2,
      "focus": "Kernel Hub availability of ops TRL could adopt." },
    { "repo": "bitsandbytes-foundation/bitsandbytes", "tier": 2,
      "focus": "Declared TRL dep, currently pinned `<0.50.0` as a temporary hotfix for TRL #6543. Watch releases and the deprecated-torch._check_is_size fix (PR #1940) so the pin can be lifted." },
    { "repo": "Dao-AILab/flash-attention", "tier": 2,
      "focus": "Attention backend releases and breaking changes that reach TRL through transformers' attn_implementation." },
    { "repo": "huggingface/tokenizers", "tier": 2,
      "focus": "Rust tokenizer releases; padding/truncation and offset-mapping changes that silently shift TRL's masking." },
    { "repo": "huggingface/math-verify", "tier": 2,
      "focus": "Declared TRL dep (`math-verify>=0.5.2`) behind the math reward functions. Low volume — releases and parsing-behaviour changes only." },
    { "repo": "harbor-framework/harbor", "tier": 2,
      "focus": "Backs the `trl[harbor]` extra and trl/experimental/harbor/_env.py sandboxed-env integration. Watch for breaking changes to the start/exec/verify/stop lifecycle TRL drives synchronously." },
    { "repo": "sgl-project/sglang", "tier": 2,
      "focus": "Alternative rollout backend features; RL-oriented weight-sync/sleep APIs that vLLM lacks." },
    { "repo": "verl-project/verl", "tier": 2,
      "focus": "Competing RL post-training library: new algorithms, scaling/perf claims, async rollout design — trend signal for what TRL users will ask for." },
    { "repo": "OpenRLHF/OpenRLHF", "tier": 2,
      "focus": "Same as verl: algorithm and infra trends in open RLHF." },
    { "repo": "THUDM/slime", "tier": 2,
      "focus": "SGLang+Megatron RL post-training framework. Watch its rollout/training disaggregation and async off-policy design — the scaling patterns TRL users will cite when asking why TRL is slower at scale." },
    { "repo": "PrimeIntellect-ai/prime-rl", "tier": 2,
      "focus": "Decentralized/async agentic RL at scale. Watch async off-policy trainer design, FSDP2 usage, and their vLLM weight-sync path — closest competitor to TRL's async GRPO direction." },
    { "repo": "PrimeIntellect-ai/verifiers", "tier": 2,
      "focus": "RL environments + evals library (formerly willccbb/verifiers; the old path redirects). The env/rubric interface is becoming a de-facto standard for RLVR — watch for it hardening into a spec TRL would need to interop with, and for reward/rubric patterns TRL users will expect from reward_funcs." },
    { "repo": "allenai/open-instruct", "tier": 2,
      "focus": "Tulu/OLMo post-training codebase, very active. Watch algorithm choices and RLVR recipes — the research-side counterweight to the infra-heavy competitors." },
    { "repo": "NVIDIA-NeMo/RL", "tier": 2,
      "focus": "NVIDIA's RL post-training stack. Watch Megatron/vLLM integration and any TRL-comparison benchmarks." },
    { "repo": "OpenPipe/ART", "tier": 2,
      "focus": "Agentic RL trainer with strong adoption. Watch its trajectory/reward interface — signal for where agentic RL ergonomics are heading." },
    { "repo": "axolotl-ai-cloud/axolotl", "tier": 2,
      "focus": "Downstream consumer of TRL — breakage reports, pins, and features they build on top of TRL trainers." },
    { "repo": "unslothai/unsloth", "tier": 2,
      "focus": "Downstream consumer of TRL — patches/workarounds they apply to TRL trainers reveal TRL bugs and perf gaps." },
    { "repo": "NVIDIA/TransformerEngine", "tier": 2,
      "focus": "FP8/attention backends reaching transformers, and thus TRL." },
    { "repo": "NVIDIA/Megatron-LM", "tier": 2,
      "focus": "Parallelism techniques (CP/SP/EP) that set expectations for TRL long-context training." },
    { "repo": "huggingface/lighteval", "tier": 2,
      "focus": "Evaluation hooks TRL users pair with training." },
    { "repo": "gradio-app/trackio", "tier": 2,
      "focus": "Experiment tracking backend used by TRL." },
    { "repo": "pytorch/torchtitan", "tier": 2,
      "focus": "Reference distributed-training patterns (FSDP2 + CP + PP) worth tracking." }
  ],

  "web": [
    { "name": "HF blog", "url": "https://huggingface.co/blog/feed.xml",
      "focus": "Post-training / RL / open-model releases that shift what TRL must support." },
    { "name": "PyTorch blog", "url": "https://pytorch.org/blog/feed.xml",
      "focus": "Distributed + compile + low-precision announcements." },
    { "name": "vLLM blog", "url": "https://api.github.com/repos/vllm-project/vllm-project.github.io/contents/_posts",
      "focus": "Engine roadmap and RL-integration posts. blog.vllm.ai has no RSS AND is JS-rendered, so WebFetch returns nothing — read the post markdown from the site's source repo instead. Filenames are YYYY-MM-DD-slug.md, so the date filter is the filename." },
    { "name": "HF papers (daily)", "url": "https://huggingface.co/api/daily_papers?limit=50",
      "focus": "New post-training / RLHF / reasoning-RL methods with real traction — candidate TRL trainers or loss variants. The /papers HTML page is JS-rendered and does not parse; this JSON API does, and carries .paper.upvotes for judging traction." }
  ],

  "pin_watch": {
    "_comment": [
      "Every run, diff TRL's declared version constraints against what upstream has actually released.",
      "Ceilings and exclusions are temporary by construction — this catches the day they become liftable,",
      "or the day a new release lands outside the tested range. Read the pins from pyproject.toml at",
      "runtime (never hardcode them here) so this keeps working as they change."
    ],
    "manifest": "/trl/pyproject.toml",
    "report_when": [
      "a release exists ABOVE an upper bound (e.g. vllm <=X and X+1 shipped) — TRL is now untested against current upstream",
      "a release exists above an exclusion/hotfix pin (e.g. bitsandbytes <0.50.0) — check whether the reason still holds",
      "the upstream issue/PR named in the pin's trailing comment has since merged or been released — the pin is liftable",
      "a floor (>=) is now far behind, such that TRL is carrying compatibility code for versions nobody runs"
    ],
    "pypi_only": [
      { "package": "openreward", "focus": "Declared TRL dep (python>=3.11) with no public source repo — watch PyPI releases only. Adapters live under the OpenRewardAI org." },
      { "package": "harbor", "focus": "Cross-check PyPI releases against harbor-framework/harbor." }
    ]
  },

  "searches": [
    { "name": "TRL used/broken elsewhere",
      "query": "\"from trl import\" OR \"trl.trainer\" OR SFTTrainer OR GRPOTrainer OR DPOTrainer created:>{SINCE_DATE}",
      "focus": "Breakage reports, integration requests, and workarounds for TRL raised in other projects. Bare 'trl' is useless as a term (acronym collisions) — these symbol names are the precise probe.",
      "note": "GitHub search IGNORES negated -repo: qualifiers. Filter out huggingface/trl and obvious junk repos client-side with jq." },
    { "name": "RFCs and roadmaps",
      "query": "label:RFC updated:>{SINCE_DATE}",
      "repos": ["vllm-project/vllm", "huggingface/transformers", "huggingface/accelerate", "pytorch/pytorch"],
      "focus": "Upcoming design changes before they ship. Run once per repo with repo: prepended — multiple repo: qualifiers in one query AND together and return nothing." }
  ]
}

Xet Storage Details

Size:: 11.3 kB
Xet hash:: c442c3e7222a37feb619fb659d66ab4bef2ff2bd9c29c59f4ceaaac8aafba0b3

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.

	{
	"_comment": [
	"Configuration for the TRL ecosystem intelligence bot.",
	"The bot runs as a Hugging Face scheduled Job. Everything it needs lives in the bucket",
	"hf://buckets/qgallouedec/ecosystem-watch, mounted read-write at /eco:",
	" sources.json this file prompt.txt the bot's instructions",
	" run.sh the Job entrypoint",
	" state/ seen.jsonl dedupe ledger + watchlist.jsonl + last_run stamp",
	" logs/ one transcript per run, named by UTC start time",
	"Edit this file to add/remove sources — the prompt reads it at runtime, nothing else needs changing.",
	"tier 1 = deep scan (releases, merged PRs, open PRs, issues, discussions, milestones).",
	"tier 2 = shallow scan (releases + notable/labelled activity only).",
	"'focus' tells the bot what kind of change in that repo actually matters to TRL."
	],

	"_slack_setup": [
	"This bot posts to its OWN channel, separate from the issue-triage bot.",
	"One-time setup, done by a human (the bot token has chat:write but not channels:manage,",
	"so it cannot create a channel or look one up by name):",
	" 1. In Slack, create the channel (suggested: #trl-ecosystem).",
	" 2. In that channel, run: /invite @trl_triage_bot",
	" Without this, chat.postMessage fails with not_in_channel.",
	" 3. Copy the channel ID (channel name > View channel details > bottom of the About tab,",
	" or the last path segment of the channel URL — it starts with C).",
	" 4. Paste it into slack_channel below, replacing REPLACE_ME.",
	"",
	"run.sh preflights this and exits without burning an API run if it is unset."
	],
	"slack_channel": "C0BLEAM73KK",
	"lookback_hours_default": 30,
	"max_items": 6,

	"github": [
	{ "repo": "huggingface/transformers", "tier": 1,
	"focus": "Trainer/TrainingArguments API changes, deprecations, model loading (dtype/device_map/quantization), attention implementations & masking, cache classes, chat templates & processors, generation config, tokenizer behaviour, anything that breaks TRL imports or silently changes loss/logits." },
	{ "repo": "huggingface/accelerate", "tier": 1,
	"focus": "FSDP2/DeepSpeed integration, gradient accumulation semantics, dataloader sharding, context parallelism, plugin/config schema changes, prepare() behaviour." },
	{ "repo": "vllm-project/vllm", "tier": 1,
	"focus": "Releases and breaking API changes in LLM/AsyncLLM, sleep/wake_up & weight-hot-swap paths, collective_rpc, V1 engine internals TRL depends on for the vLLM backend, logprobs semantics, server/OpenAI-API changes." },
	{ "repo": "huggingface/peft", "tier": 1,
	"focus": "LoRA/adapter API changes, merge/unmerge and adapter hot-swap, new PEFT methods worth exposing in TRL configs." },
	{ "repo": "huggingface/datasets", "tier": 1,
	"focus": "Dataset/IterableDataset API changes, map/column semantics, streaming, anything affecting TRL dataset formats or collators." },
	{ "repo": "huggingface/huggingface_hub", "tier": 1,
	"focus": "Hub upload/push_to_hub, model cards, inference/jobs API, auth changes used by TRL." },

	{ "repo": "deepspeedai/DeepSpeed", "tier": 2,
	"focus": "ZeRO-3 gather/partition behaviour, memory regressions, compatibility with recent transformers/accelerate." },
	{ "repo": "pytorch/pytorch", "tier": 2,
	"focus": "Release notes and deprecations touching torch.compile, FSDP2, DTensor, autocast/dtype APIs, distributed primitives." },
	{ "repo": "pytorch/ao", "tier": 2,
	"focus": "Quantized/low-precision training (QAT, fp8, NVFP4) usable from TRL." },
	{ "repo": "linkedin/Liger-Kernel", "tier": 2,
	"focus": "Fused loss kernels (chunked CE, DPO/GRPO losses) TRL wires up, and their API/signature changes." },
	{ "repo": "huggingface/kernels", "tier": 2,
	"focus": "Kernel Hub availability of ops TRL could adopt." },
	{ "repo": "bitsandbytes-foundation/bitsandbytes", "tier": 2,
	"focus": "Declared TRL dep, currently pinned `<0.50.0` as a temporary hotfix for TRL #6543. Watch releases and the deprecated-torch._check_is_size fix (PR #1940) so the pin can be lifted." },
	{ "repo": "Dao-AILab/flash-attention", "tier": 2,
	"focus": "Attention backend releases and breaking changes that reach TRL through transformers' attn_implementation." },
	{ "repo": "huggingface/tokenizers", "tier": 2,
	"focus": "Rust tokenizer releases; padding/truncation and offset-mapping changes that silently shift TRL's masking." },
	{ "repo": "huggingface/math-verify", "tier": 2,
	"focus": "Declared TRL dep (`math-verify>=0.5.2`) behind the math reward functions. Low volume — releases and parsing-behaviour changes only." },
	{ "repo": "harbor-framework/harbor", "tier": 2,
	"focus": "Backs the `trl[harbor]` extra and trl/experimental/harbor/_env.py sandboxed-env integration. Watch for breaking changes to the start/exec/verify/stop lifecycle TRL drives synchronously." },
	{ "repo": "sgl-project/sglang", "tier": 2,
	"focus": "Alternative rollout backend features; RL-oriented weight-sync/sleep APIs that vLLM lacks." },
	{ "repo": "verl-project/verl", "tier": 2,
	"focus": "Competing RL post-training library: new algorithms, scaling/perf claims, async rollout design — trend signal for what TRL users will ask for." },
	{ "repo": "OpenRLHF/OpenRLHF", "tier": 2,
	"focus": "Same as verl: algorithm and infra trends in open RLHF." },
	{ "repo": "THUDM/slime", "tier": 2,
	"focus": "SGLang+Megatron RL post-training framework. Watch its rollout/training disaggregation and async off-policy design — the scaling patterns TRL users will cite when asking why TRL is slower at scale." },
	{ "repo": "PrimeIntellect-ai/prime-rl", "tier": 2,
	"focus": "Decentralized/async agentic RL at scale. Watch async off-policy trainer design, FSDP2 usage, and their vLLM weight-sync path — closest competitor to TRL's async GRPO direction." },
	{ "repo": "PrimeIntellect-ai/verifiers", "tier": 2,
	"focus": "RL environments + evals library (formerly willccbb/verifiers; the old path redirects). The env/rubric interface is becoming a de-facto standard for RLVR — watch for it hardening into a spec TRL would need to interop with, and for reward/rubric patterns TRL users will expect from reward_funcs." },
	{ "repo": "allenai/open-instruct", "tier": 2,
	"focus": "Tulu/OLMo post-training codebase, very active. Watch algorithm choices and RLVR recipes — the research-side counterweight to the infra-heavy competitors." },
	{ "repo": "NVIDIA-NeMo/RL", "tier": 2,
	"focus": "NVIDIA's RL post-training stack. Watch Megatron/vLLM integration and any TRL-comparison benchmarks." },
	{ "repo": "OpenPipe/ART", "tier": 2,
	"focus": "Agentic RL trainer with strong adoption. Watch its trajectory/reward interface — signal for where agentic RL ergonomics are heading." },
	{ "repo": "axolotl-ai-cloud/axolotl", "tier": 2,
	"focus": "Downstream consumer of TRL — breakage reports, pins, and features they build on top of TRL trainers." },
	{ "repo": "unslothai/unsloth", "tier": 2,
	"focus": "Downstream consumer of TRL — patches/workarounds they apply to TRL trainers reveal TRL bugs and perf gaps." },
	{ "repo": "NVIDIA/TransformerEngine", "tier": 2,
	"focus": "FP8/attention backends reaching transformers, and thus TRL." },
	{ "repo": "NVIDIA/Megatron-LM", "tier": 2,
	"focus": "Parallelism techniques (CP/SP/EP) that set expectations for TRL long-context training." },
	{ "repo": "huggingface/lighteval", "tier": 2,
	"focus": "Evaluation hooks TRL users pair with training." },
	{ "repo": "gradio-app/trackio", "tier": 2,
	"focus": "Experiment tracking backend used by TRL." },
	{ "repo": "pytorch/torchtitan", "tier": 2,
	"focus": "Reference distributed-training patterns (FSDP2 + CP + PP) worth tracking." }
	],

	"web": [
	{ "name": "HF blog", "url": "https://huggingface.co/blog/feed.xml",
	"focus": "Post-training / RL / open-model releases that shift what TRL must support." },
	{ "name": "PyTorch blog", "url": "https://pytorch.org/blog/feed.xml",
	"focus": "Distributed + compile + low-precision announcements." },
	{ "name": "vLLM blog", "url": "https://api.github.com/repos/vllm-project/vllm-project.github.io/contents/_posts",
	"focus": "Engine roadmap and RL-integration posts. blog.vllm.ai has no RSS AND is JS-rendered, so WebFetch returns nothing — read the post markdown from the site's source repo instead. Filenames are YYYY-MM-DD-slug.md, so the date filter is the filename." },
	{ "name": "HF papers (daily)", "url": "https://huggingface.co/api/daily_papers?limit=50",
	"focus": "New post-training / RLHF / reasoning-RL methods with real traction — candidate TRL trainers or loss variants. The /papers HTML page is JS-rendered and does not parse; this JSON API does, and carries .paper.upvotes for judging traction." }
	],

	"pin_watch": {
	"_comment": [
	"Every run, diff TRL's declared version constraints against what upstream has actually released.",
	"Ceilings and exclusions are temporary by construction — this catches the day they become liftable,",
	"or the day a new release lands outside the tested range. Read the pins from pyproject.toml at",
	"runtime (never hardcode them here) so this keeps working as they change."
	],
	"manifest": "/trl/pyproject.toml",
	"report_when": [
	"a release exists ABOVE an upper bound (e.g. vllm <=X and X+1 shipped) — TRL is now untested against current upstream",
	"a release exists above an exclusion/hotfix pin (e.g. bitsandbytes <0.50.0) — check whether the reason still holds",
	"the upstream issue/PR named in the pin's trailing comment has since merged or been released — the pin is liftable",
	"a floor (>=) is now far behind, such that TRL is carrying compatibility code for versions nobody runs"
	],
	"pypi_only": [
	{ "package": "openreward", "focus": "Declared TRL dep (python>=3.11) with no public source repo — watch PyPI releases only. Adapters live under the OpenRewardAI org." },
	{ "package": "harbor", "focus": "Cross-check PyPI releases against harbor-framework/harbor." }
	]
	},

	"searches": [
	{ "name": "TRL used/broken elsewhere",
	"query": "\"from trl import\" OR \"trl.trainer\" OR SFTTrainer OR GRPOTrainer OR DPOTrainer created:>{SINCE_DATE}",
	"focus": "Breakage reports, integration requests, and workarounds for TRL raised in other projects. Bare 'trl' is useless as a term (acronym collisions) — these symbol names are the precise probe.",
	"note": "GitHub search IGNORES negated -repo: qualifiers. Filter out huggingface/trl and obvious junk repos client-side with jq." },
	{ "name": "RFCs and roadmaps",
	"query": "label:RFC updated:>{SINCE_DATE}",
	"repos": ["vllm-project/vllm", "huggingface/transformers", "huggingface/accelerate", "pytorch/pytorch"],
	"focus": "Upcoming design changes before they ship. Run once per repo with repo: prepended — multiple repo: qualifiers in one query AND together and return nothing." }
	]
	}