{"full_name":"oumi-ai/oumi","owner":"oumi-ai","name":"oumi","shard":"1","url":"https://github.com/oumi-ai/oumi","description":"Easily fine-tune (SFT/RL), evaluate, and deploy Qwen, Gemma, or any open agentic LLM/VLM!","homepage":"https://oumi.ai","topics":["dpo","fine-tuning","inference","llama","llms","sft","gpt-oss","agentic-ai","gemma","grpo","llm","llm-as-a-judge","llm-evaluation","lora","pytorch"],"tag_pages":["agentic-ai","fine-tuning","gemma","gpt-oss","grpo","inference","llama","llm","llm-evaluation","llms","lora","pytorch"],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/167452922?v=4","created_at":"2024-05-07T17:45:15Z","pushed_at":"2026-11-12T17:36:38Z","metrics":{"stars":9433,"forks":798,"watchers":61,"issues":117,"prs":2606,"releases":27,"issues_open":1,"issues_closed":116,"prs_open":25,"prs_closed":241,"prs_merged":2340,"discussions":0},"growth":{"day":{"stars":0,"forks":0,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[9430,9430,9430,9430,9431,9433,9433],"forks":[798,798,798,798,798,798,798],"watchers":[null,61,61,61,61,61,61],"issues_open":[null,1,1,1,1,1,1],"prs_open":[null,24,22,20,25,25,25]},"boards":{},"same_owner":[],"issues":[{"number":2730,"title":"[Feature] Nonblocking execution for synchronous VERL tools","comments":0,"created_at":"2026-10-29T10:49:25Z"}],"discussions":[],"latest_release":{"name":"v0.9","tag":"v0.9","published_at":"2026-09-16T21:48:48Z","body":"# Oumi v0.9 Release Notes\r\n\r\nThis release makes Oumi an end-to-end stack for **agentic, tool-using models**: you can now train on tool-calling data, execute tool calls against real environments (databases, HTTP endpoints, deterministic lookups, and model-simulated tools), synthesize multi-turn tool conversations, and run GRPO/RL over any of those environments with verl. It also adds a multi-criteria `RubricJudge`, partial-failure-tolerant inference/synthesis/judging, a **Modal** GPU launcher and **SkyPilot-routed Slurm**, and new **Gemma 4** recipes.\r\n\r\n## Highlights\r\n\r\n### 🛠️ Tool-use training\r\n\r\nTrain models to call tools end-to-end:\r\n- Tool-aware chat templates — access the tool-call section, auto-detect the tool-call form a template expects *(#2605, #2603, #2604)*\r\n- Trainable tool adapters *(#2658)*\r\n- Collator masking that correctly remasks tool-response spans, including responses nested in an assistant turn *(#2594)*\r\n- Tool-calling SFT and text DPO over JSONL, with serialized tool arguments decoded during tokenization *(#2642, #2517)*\r\n- Canonical tool-call `Conversation` shape + JSON-string ↔ dict argument helpers *(#2602, #2601)*\r\n\r\n### 🌐 Tool-execution Environments\r\n\r\nThe new `oumi.environments` subsystem supports executing tools calls with various backends.\r\n\r\n- **`ExecutableEnvironment` / `ExecutableTool`** — the base abstraction for tools that actually run. *(#2527)*\r\n- **`DatabaseExecutableEnvironment`** — run tool calls against a database session, with rollback-isolated teardown (see the EHR example). *(#2528, #2545)*\r\n- **`EndpointEnvironment`** — dispatch tool calls to HTTP endpoints. *(#2625)*\r\n- **Lookup (deterministic) and simulated environments** — deterministic lookup tables validated against tool schemas, with argument defaults applied and any JSON value as output; model-simulated tool execution for synthesis. *(#2623, #2536, #2573, #2574)*\r\n- Tool executors are resolvable by registry name, so environments can wire tools declaratively. *(#2555)*\r\n\r\n### 🎯 `OumiVerlTool` — RL over any Oumi environment\r\n\r\n`OumiVerlTool` bridges Oumi's executable environments into verl, so you can run verl GRPO tool-use RL over any Oumi environment. Custom judge rewards are now supported in verl GRPO, multi-turn prompts are supported in the GRPO data path, conversation metadata is preserved through GRPO datasets, and per-run verl metrics are mirrored to `verl_metrics.jsonl` with richer logging.\r\n*(#2569, #2624, #2542, #2644, #2643, #2646)*\r\n\r\n### ⚖️ `RubricJudge` — multi-criteria evaluation\r\n\r\nA new `RubricJudge` scores a response against several named criteria in a single judge call, instead of one boolean per pass — useful both for evaluation and as a GRPO reward. Ships with example configs under `configs/projects/judges/rubric` and documentation.\r\n*(#2626, #2628, #2629)*\r\n\r\n### 🧬 Conversation synthesis\r\n\r\nSynthesize multi-turn tool-use conversations:\r\n- Stateful and stateless synthetic tool environments for multi-turn tool interactions *(#2410, #2457)*\r\n- Bring your own inference engine/provider for synthesis *(#2504)*\r\n- Reusable `ConversationSynthesizer` building blocks to drive synthesis programmatically *(#2531)*\r\n- Token-usage accounting across a synthesis run *(#2480)*\r\n- Misc improvements: #2559, #2630, #2588, #2591, #2576\r\n\r\n### 🔁 Partial-failure-tolerant pipelines\r\n\r\nLong inference, synthesis, and judging runs can now return partial results instead of failing the whole batch. New `infer_partial()` / `synthesize_partial()` / `judge_partial()` templates, per-row partial failures in `RemoteInferenceEngine`, partial-inference result types, and a progress-file reporter you can poll while a run is in flight.\r\n*(#2498, #2499, #2500, #2501, #2502)*\r\n\r\n### ⚡ Inference hardening\r\n\r\n- **Anthropic**: configurable prompt-cache TTL, cache tokens folded into `prompt_tokens`, correct model-version gating for round and dated model names, and sampling params omitted for models that reject them. *(#2665, #2620, #2568, #2538)*\r\n- *","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}