Reinforcement Learning
Transformers
English
post-training
distillation
agentic-coding
composer-2.5
cursor
kimi-k2
grpo
dapo
diloco
openenv
trl
verl
research
methodology
Instructions to use Codeseys/composer-replication-framework with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Codeseys/composer-replication-framework with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("Codeseys/composer-replication-framework", dtype="auto") - Notebooks
- Google Colab
- Kaggle
| {"applied": [{"type": "run-on", "where": "§2 (line 27, killer-fact sentence)", "what": "Split semicolon-joined [11]/[55] run-on (~75 words) into two sentences at 'In SWE specifically'."}, {"type": "run-on", "where": "§2 (line 27, MuZero/Dreamer)", "what": "Split '; and the latent-motion line' into '. The latent-motion line', breaking the two-discipline run-on."}, {"type": "run-on", "where": "§2 (line 29, Predictive-Causal Gap)", "what": "Split 'decision-relevant; the value-equivalent target' into two sentences."}, {"type": "run-on", "where": "§2 (line 33, SDPO carrier)", "what": "Split em-dash run-on before 'ADR-011's placeholder-system-message' into two sentences (~55 words)."}, {"type": "run-on", "where": "§2 (line 35, Measurement)", "what": "Split triple-semicolon run-on so the foresight@k kill-ablation definition stands alone."}, {"type": "redundancy", "where": "§4 (line 75, two-harvest frame)", "what": "Removed meta-sentence restating 'two independent lines of analysis converge on one mechanism' — already delivered in Opinionated Synthesis ('Three roles, one lever'); the bolded sentence + parenthetical fully carry the point in-section."}, {"type": "run-on", "where": "§4 (line 79, hack-surface qualifier)", "what": "Broke ~150-word multi-semicolon run-on: separated the EvilGenie clause and the safeguard-#1 conclusion into discrete sentences; preserved [30][29][31][1] and all claims."}, {"type": "run-on", "where": "§8 (line 179, hosting fact)", "what": "Split '(PRIME-RL too); and TRL has no async' into two sentences."}, {"type": "run-on", "where": "§8 (line 179, layered sandbox posture)", "what": "Broke the third isolation tier (container-free SWE-MiniSandbox) off the ~85-word tri-tier run-on into its own sentence; kept gVisor/Kata parallelism."}, {"type": "run-on", "where": "§5 (line 103, self-distillation stabilizer)", "what": "Split stacked-em-dash run-on (~95 words) at '— though the repo's own ADR-013' into two sentences."}, {"type": "run-on", "where": "§5 (line 103, flywheel)", "what": "Split 'OOD [37]); most collapse stories' at the clean clause boundary into two sentences; preserved [10][37][43][29][38]."}], "escalations": []} | |