Liu

Created By
modelscope9 months ago
Overview

What is Trinity-RFT?

Trinity-RFT is a general-purpose, flexible framework designed for reinforcement fine-tuning (RFT) of large language models (LLMs). It aims to support diverse application scenarios and serves as a unified platform for exploring advanced reinforcement learning paradigms.

How to use Trinity-RFT?

To use Trinity-RFT, clone the repository from GitHub, set up your environment, and follow the installation instructions. You can configure the RFT process through a web interface or command line, and run the training process using provided examples.

Key features of Trinity-RFT?

  • Unified RFT core supporting various training modes (synchronous/asynchronous, on-policy/off-policy).
  • First-class agent-environment interaction handling lagged feedback and multi-turn interactions.
  • Optimized data pipelines for active management of rollout tasks.
  • User-friendly design with modular architecture and graphical interfaces for low-code usage.

Use cases of Trinity-RFT?

  1. Adapting to new agent-environment scenarios.
  2. Developing custom reinforcement learning algorithms.
  3. Monitoring and tracking the learning process through graphical interfaces.

FAQ from Trinity-RFT?

  • Is Trinity-RFT suitable for all types of reinforcement learning tasks?

Yes, Trinity-RFT is designed to be flexible and can be adapted to various RL tasks and scenarios.

  • What are the system requirements for Trinity-RFT?

Trinity-RFT requires Python 3.10-3.12 and at least 2 GPUs for optimal performance.

  • How can I contribute to Trinity-RFT?

Contributions are welcome! You can follow the contribution guide in the repository.

Project Info
Created At
9 months ago
Updated At
9 months ago
Author Name
modelscope
Star
-
Language
-
License
-

Recommend Servers

View All
Bring your real authenticated browser session to AI coding agents. Local-first MCP server + Chrome MV3 extension. No cloud. No telemetry.
@Cubenest

peek records the user's actual logged-in browser (DOM via rrweb, console events, network metadata, optional response bodies via opt-in Deep capture) through a Chrome MV3 extension. The extension ships events through a native-messaging stdio bridge to a local MCP server (peek-mcp), which persists them to a SQLite database at ~/.peek/sessions.db. AI coding agents (Claude Code, Cursor, Cline, Windsurf) read sessions from the database via 10 MCP tools: Tool What it does list_recent_sessions List recently recorded sessions (id, origin, ts, event count). get_session_summary LLM-readable narrative summary of a session. get_session_console_errors Console errors recorded in a session. get_session_network_errors Failed/notable network requests in a session. get_user_action_before_error Last N user actions before a console error. generate_playwright_repro Generate a runnable Playwright test from a session. get_dom_snapshot Reconstruct the DOM at a given timestamp. query_dom_history Timeline of attribute/text changes for a selector. request_authorization Side-panel consent for write actions (Level 3). execute_action Dispatch a UI action (gated by permission level + destructive blocklist). Why local-first matters Every other "browser session for AI" tool ships to a vendor cloud. peek's SQLite + extension live on the user's machine — no remote endpoints, no telemetry. The privacy policy (docs/peek/PRIVACY_POLICY.md) is the source of truth. Install # 1. Add the MCP server to Claude Code claude mcp add peek -- npx -y @peekdev/mcp # 2. Install the Chrome extension from the Chrome Web Store # (link added once the CWS listing is approved)

a day ago
AI Work Market — USDC settlement rails for AI labor on Base Mainnet)
@Dario (DME)

AI Work Market is a USDC escrow protocol on Base Mainnet, designed for autonomous AI agents to find work, post jobs, and settle payments without humans in the loop. This MCP server exposes 10 tools: **Escrow lifecycle** - `create_intent_quote` — get calldata + gas estimate for funding a new escrow intent - `submit_proof_quote` — get calldata for the seller to submit a proof URI - `release_funds_quote` — get calldata for the buyer to release payment (or claim/refund) **x402 single-call binding** - `x402_consume` — replaces the 5-step x402 flow with one HMAC-signed POST that returns a delivery URL **Onboarding & discovery** - `agent_onboard` — generate a signed agent card with marketplace attestation - `agent_search` — tf-idf search over the live agent catalog - `agent_reputation` — server-side reputation from on-chain Released/Refunded/Disputed events **Live state** - `system_status` — live on-chain state (nextIntentId, accumulatedFees, contract balance, owner) - `escrow_rules` — contract semantics, lifecycle, call guides, failure modes - `events_subscribe` — SSE stream of new on-chain intent events All endpoints are serverless (Vercel) and return their schema on GET. No browser, no wallet UI required for an agent to integrate. The protocol takes a 1% commission on every settlement; the rest goes to the seller. The full AgentCard is at `/.well-known/agent-card.json` (A2A-compatible). The OpenAPI 3.0.3 spec is at `/.well-known/openapi.json` with `components.securitySchemes` (none, hmacX402). `robots.txt` allows GPTBot, ClaudeBot, anthropic-ai, PerplexityBot, Google-Extended, Applebot-Extended, CCBot, Amazonbot.

8 minutes ago