NousResearch/atropos
1
0
Fork 0
mirror of https://github.com/NousResearch/atropos.git synced 2026-04-19 12:57:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Update README.md

Browse source
This commit is contained in:
VolodymyrBg 2026-01-29 10:23:53 +02:00 • committed by GitHub
parent dd02df0d76
commit e0744adf28
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
atroposlib/envs/README.md
View file

@ -15,7 +15,7 @@ To achieve this generality, our environment abstraction deviates from other open
- **Groups as atomic units of data**: A natural choice for a data atom in RL is a single trajectory. However, many popular RL methods for fine-tuning LLMs such as DPO and GRPO involve packing contrastive data into the same batch. As such the most fundamental dataloading method in our abstraction is not `collect_trajectory` (singular) but `collect_trajectories` (plural). We do not enforce any definition of what a "group" is other than a set of rollouts. Although a "group" is most commonly constructed by generated multiple rollouts starting from the same initial state (as in DPO and GRPO), a user could just as easily pack `n` similar-sounding problems with very different solutions into a group. For cases like PPO where advantages don't depend on group statistics a user can simply use group size 1.
- **Environments return tokens (not messages!)**: One of the most peculiar design choices we made was that at least for text-only environments, environments are responsible for tokenization. This gives us the flexibility to assign token-level rewards and to mix completions-based (e.g. autocomplete suggestion accept/reject) and chat-based (e.g. instruct-model code generation) environments together in the same training run. For cases like multimodal where a OpenAI-formatted message list needs to be passed to a transformers `AutoProcessor`, we support a `list[dict]`-valued `messages` key within our group abstraction [ScoredDataGroup](https://github.com/NousResearch/atropos/blob/a282604baac8dbb3b201f992cfc889ee1e5a0f4a/atroposlib/envs/base.py#L55).
- **Environments return tokens (not messages!)**: One of the most peculiar design choices we made was that at least for text-only environments, environments are responsible for tokenization. This gives us the flexibility to assign token-level rewards and to mix completions-based (e.g. autocomplete suggestion accept/reject) and chat-based (e.g. instruct-model code generation) environments together in the same training run. For cases like multimodal where an OpenAI-formatted message list needs to be passed to a transformers `AutoProcessor`, we support a `list[dict]`-valued `messages` key within our group abstraction [ScoredDataGroup](https://github.com/NousResearch/atropos/blob/a282604baac8dbb3b201f992cfc889ee1e5a0f4a/atroposlib/envs/base.py#L55).
## Working with Servers and ManagedServer