NousResearch/atropos
1
0
Fork 0
mirror of https://github.com/NousResearch/atropos.git synced 2026-04-22 16:48:57 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

narrow down scope further

Browse source
This commit is contained in:
Jai Suphavadeeprasit 2026-02-27 13:15:23 -05:00
parent f343b24a6a
commit 836c346406
3 changed files with 22 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

10
example_trainer/README.md
View file

@ -8,6 +8,16 @@ This example uses `vLLM` for efficient inference during the (simulated) data gen
**Note:** This script is intended as a *reference example* for API integration and basic training setup. It is not optimized for large-scale, efficient training.
## On-Policy Distillation Scope
The current OPD integration in Atropos is transport-only:
- `ScoredDataGroup` / API payloads support `distill_token_ids` and `distill_logprobs`.
- Atropos API stores and returns those fields through `/scored_data` and `/batch`.
- Teacher orchestration (teacher endpoint calls, prompt rendering, top-k fetching) is intentionally out of scope in this PR.
If you train with distillation, provide the two distill arrays from your environment or external data pipeline before posting to the API.
### Custom vLLM Server
The `vllm_api_server.py` file in this directory provides a customized vLLM API server implementation based on vLLM's native API. This server exposes enhanced endpoints for token and logprob tracking. The `VLLMServer` class in `atroposlib/envs/server_handling/vllm_server.py` can connect to this server for direct access to vLLM's `/generate` endpoint with full token-level logprobs.