Commit Graph

281 Commits (e84437a8caea62f44974ea8447bb67eff43f432c)

Author SHA1 Message Date
vanking e84437a8ca
[add] Create train_gated_grpo.py
add gating term on grpo algorithm
2026-02-03 10:20:56 +08:00
vanking c540ea2537
[add] created train_gated_ppo.py
modify ppo algorithm by adding gating term on importance ratio
2026-02-03 10:17:56 +08:00
vanking ce9ed24dcd
[add] add CISPO (Clipped Importance Sampling Policy Optimization) algorithm
This the main addition reference GRPO algorithm: By changing the gradient of out-of-bounds tokens from "directly set to 0" to "bounded clipping", we ensure that high-value exploration tokens can continue to participate in parameter updates while maintaining training stability.
Actually, this change is not very well in experiment, because the ratio is almost nearly at 1, which means seldom out-of-bound.
2026-02-02 22:23:27 +08:00
vanking 7389f64dee
[add] add DAPO argorithm (Decoupled Clip and Dynamic sAmpling Policy Optimization)
I mainly add two points  into original grpo algorithm according to this [paper](https://arxiv.org/pdf/2503.14476): Clip-Higher & Dynamic Sampling
2026-02-02 22:16:52 +08:00
vanking 35fe1399c8
Update README.md 2026-02-02 16:33:02 +08:00
Your Name 020bd44f3f [mod] fix spo algorithm in RLAIF part 2026-01-30 11:03:35 +08:00
jingyaogong a9c56b20e9 [fix] lora weight 2025-12-22 21:27:29 +08:00
jingyaogong 048d84abc7
Merge pull request #594 from whiteswordLI/fix/lora-load-ddp-weights
Fix: support loading DDP-saved LoRA weights for inference
2025-12-22 21:19:16 +08:00
whitesword 3a18fdd666 Fix: support loading DDP-saved LoRA weights for inference 2025-12-22 20:50:25 +08:00
jingyaogong fe24501602 [feat] adjust seq length 2025-12-14 20:41:58 +08:00
jingyaogong fa82707c9c [feat] update readme 2025-12-11 15:45:50 +08:00
jingyaogong 5129f0e2a2 [fix] dtype & lr 2025-12-09 13:01:38 +08:00
jingyaogong aa7dc0f61e
Merge pull request #571 from dyhuachi/dyhuachi-patch-1
[fix] Refactor get_lr function to include min_lr calculation
2025-12-09 12:59:11 +08:00
dyhuachi bf3878ace8
[fix] Refactor get_lr function to include min_lr calculation
这里的退火算法会让参数里的lr的起始值变成原来lr的1.1倍,作出如下修改
2025-12-06 17:09:51 +08:00
jingyaogong ecd1ae1563 [fix] reduce aux_loss_alpha 2025-12-05 23:08:29 +08:00
jingyaogong 5e1447b913 [fix] cuda memory #559 2025-12-01 16:17:43 +08:00
jingyaogong 151fdf7e76 [feat] update yarn 2025-12-01 16:15:05 +08:00
jingyaogong 6b86ea399a [feat] release memory 2025-11-27 19:39:49 +08:00
jingyaogong d7f4f4eab8 [fix] ppo mask 2025-11-19 23:39:02 +08:00
jingyaogong f5374dc87f [fix] model attn_mask 2025-11-19 22:26:53 +08:00
jingyaogong a044578d73 [fix] update model 2025-11-18 13:07:20 +08:00
jingyaogong ce9394670b
Merge pull request #536 from yuyu5333/fix/attn_forward
fix: attn_forwad when is_causal=True assert attn_mask is None
2025-11-18 13:02:46 +08:00
yuyu5333 7d02ce673c fix: attn_forwad when is_causal=True assert attn_mask is None 2025-11-18 03:17:17 +00:00
jingyaogong 9c98cabc9a [fix] prompt length calculate 2025-11-15 18:25:37 +08:00
jingyaogong f3441b0078
Merge pull request #528 from wangzhaode/feat/add_mnn_support
[feat] add MNN support to README.
2025-11-10 22:46:15 +08:00
yanxing 5959396096 [feat] add MNN support to README. 2025-11-10 21:59:22 +08:00
jingyaogong bf60bde8fb [fix] model-name 2025-11-07 19:38:20 +08:00
jingyaogong 81e869fc3e [fix] harmonize template 2025-11-06 13:14:08 +08:00
jingyaogong 509d8dacf1 [feat] clear cache 2025-11-06 13:12:28 +08:00
jingyaogong 8a0b04ed82 [fix] harmonize template 2025-11-02 23:18:11 +08:00
jingyaogong 0323815729 [feat] update import 2025-10-31 23:45:55 +08:00
jingyaogong 8d71754e05 [feat] update readme 2025-10-30 23:39:25 +08:00
jingyaogong d8ac558ce2 [feat] update readme 2025-10-30 23:30:12 +08:00
jingyaogong e4807a5214 [feat] update readme 2025-10-30 23:27:15 +08:00
jingyaogong de23e1ea39 [feat] update datasets 2025-10-30 11:08:13 +08:00
jingyaogong 08ce3da228 [feat] update args 2025-10-30 10:48:31 +08:00
jingyaogong bf123b585d [feat] add args 2025-10-30 10:05:12 +08:00
jingyaogong 800fed4639 [feat] update readme 2025-10-29 12:13:25 +08:00
jingyaogong 1713c24114 [fix] model device 2025-10-29 10:36:28 +08:00
jingyaogong acd5925193 [feat] update trainer 2025-10-29 00:52:37 +08:00
jingyaogong 8f7e07b8ef [feat] update trainer 2025-10-28 23:30:10 +08:00
jingyaogong eb96113cd4 [feat] update readme 2025-10-27 16:31:40 +08:00
jingyaogong cc81925ecc [feat] update readme 2025-10-26 18:52:50 +08:00
jingyaogong 35bed8936d [feat] update eval-llm 2025-10-26 18:52:01 +08:00
jingyaogong e8484874f5 [feat] pause-training 2025-10-26 18:49:52 +08:00
jingyaogong 6efba3249a [feat] update readme 2025-10-24 01:18:33 +08:00
jingyaogong ea2abb5fb3 [feat] update readme 2025-10-24 00:45:11 +08:00
jingyaogong c4e9789d7e [feat] update requirements 2025-10-23 23:54:55 +08:00
jingyaogong 4276f8a72f [feat] update requirements 2025-10-23 23:50:21 +08:00
jingyaogong d762926f48 [feat] update readme 2025-10-23 23:23:52 +08:00