amir-ammar commited on
Commit
5839dfc
·
verified ·
1 Parent(s): f555f17

flab general_knowledge/sft 20260525_104615

Browse files
README.md CHANGED
@@ -1,16 +1,17 @@
1
  ---
 
2
  library_name: transformers
3
- model_name: '20260520_190659'
4
  tags:
5
  - generated_from_trainer
 
6
  - trl
7
- - grpo
8
  licence: license
9
  ---
10
 
11
- # Model Card for 20260520_190659
12
 
13
- This model is a fine-tuned version of [None](https://huggingface.co/None).
14
  It has been trained using [TRL](https://github.com/huggingface/trl).
15
 
16
  ## Quick start
@@ -26,11 +27,11 @@ print(output["generated_text"])
26
 
27
  ## Training procedure
28
 
29
- [<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/cs-552-2026-flab/flab-cs552/runs/5txs4jcy)
30
 
31
 
32
 
33
- This model was trained with GRPO, a method introduced in [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://huggingface.co/papers/2402.03300).
34
 
35
  ### Framework versions
36
 
@@ -42,16 +43,7 @@ This model was trained with GRPO, a method introduced in [DeepSeekMath: Pushing
42
 
43
  ## Citations
44
 
45
- Cite GRPO as:
46
 
47
- ```bibtex
48
- @article{shao2024deepseekmath,
49
- title = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
50
- author = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
51
- year = 2024,
52
- eprint = {arXiv:2402.03300},
53
- }
54
- ```
55
 
56
  Cite TRL as:
57
 
 
1
  ---
2
+ base_model: Qwen/Qwen3-1.7B
3
  library_name: transformers
4
+ model_name: '20260519_023851'
5
  tags:
6
  - generated_from_trainer
7
+ - sft
8
  - trl
 
9
  licence: license
10
  ---
11
 
12
+ # Model Card for 20260519_023851
13
 
14
+ This model is a fine-tuned version of [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B).
15
  It has been trained using [TRL](https://github.com/huggingface/trl).
16
 
17
  ## Quick start
 
27
 
28
  ## Training procedure
29
 
30
+ [<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/cs-552-2026-flab/flab-cs552/runs/slclweh8)
31
 
32
 
33
 
34
+ This model was trained with SFT.
35
 
36
  ### Framework versions
37
 
 
43
 
44
  ## Citations
45
 
 
46
 
 
 
 
 
 
 
 
 
47
 
48
  Cite TRL as:
49
 
hydra/.hydra/config.yaml CHANGED
@@ -5,40 +5,40 @@ model:
5
  attn_impl: flash_attention_2
6
  max_model_len: 4096
7
  data:
8
- name: general_knowledge_rlvr
9
- loader: load_general_knowledge_rlvr
10
  kwargs:
11
- n_mmlu: null
 
 
12
  n_arc: null
 
 
 
 
 
13
  seed: 0
14
  training:
15
- recipe: rlvr
16
- num_generations: 8
17
- max_completion_length: 256
18
- temperature: 1.0
19
- top_p: 0.9
20
- beta: 0.0
21
  epochs: 1
22
  per_device_train_batch_size: 1
23
- gradient_accumulation_steps: 8
24
- learning_rate: 1.0e-06
25
- warmup_ratio: 0.05
26
  lr_scheduler_type: cosine
27
  gradient_checkpointing: true
28
  bf16: true
29
- logging_steps: 5
30
- save_steps: 50
31
- save_total_limit: 2
32
  max_seq_length: 4096
33
- deepspeed: null
 
 
34
  report_to: wandb
35
- use_vllm: false
36
  domain:
37
  name: general_knowledge
38
  system_instruction: null
39
  thinking_mode: true
40
  benchmark: knowledge
41
- chat_template_path: null
42
  seed: 42
43
  run_id: ${now:%Y%m%d_%H%M%S}
44
  output_dir: ${oc.env:FLAB_SCRATCH,/scratch}/checkpoints/${domain.name}/${training.recipe}/${run_id}
@@ -62,7 +62,7 @@ es:
62
  base_model_dir: null
63
  data: null
64
  rlvr:
65
- base_model_dir: /scratch/checkpoints/general_knowledge/sft/20260519_023851
66
  dry_run: false
67
  hf_push:
68
  enabled: false
 
5
  attn_impl: flash_attention_2
6
  max_model_len: 4096
7
  data:
8
+ name: general_knowledge_bare_mc_sft
9
+ loader: load_general_knowledge_sft
10
  kwargs:
11
+ bare_letter: true
12
+ n_mmlu_pro: null
13
+ n_mmlu: 25000
14
  n_arc: null
15
+ n_medmcqa: 5000
16
+ n_gpqa: null
17
+ n_sciq: 0
18
+ n_scienceqa: 0
19
+ max_total_tokens: 3500
20
  seed: 0
21
  training:
22
+ recipe: sft
 
 
 
 
 
23
  epochs: 1
24
  per_device_train_batch_size: 1
25
+ gradient_accumulation_steps: 16
26
+ learning_rate: 5.0e-06
27
+ warmup_ratio: 0.03
28
  lr_scheduler_type: cosine
29
  gradient_checkpointing: true
30
  bf16: true
 
 
 
31
  max_seq_length: 4096
32
+ logging_steps: 10
33
+ save_steps: 500
34
+ save_total_limit: 2
35
  report_to: wandb
36
+ deepspeed: null
37
  domain:
38
  name: general_knowledge
39
  system_instruction: null
40
  thinking_mode: true
41
  benchmark: knowledge
 
42
  seed: 42
43
  run_id: ${now:%Y%m%d_%H%M%S}
44
  output_dir: ${oc.env:FLAB_SCRATCH,/scratch}/checkpoints/${domain.name}/${training.recipe}/${run_id}
 
62
  base_model_dir: null
63
  data: null
64
  rlvr:
65
+ base_model_dir: null
66
  dry_run: false
67
  hf_push:
68
  enabled: false
hydra/.hydra/hydra.yaml CHANGED
@@ -112,20 +112,14 @@ hydra:
112
  hydra:
113
  - hydra.mode=RUN
114
  task:
115
- - rlvr.base_model_dir=/scratch/checkpoints/general_knowledge/sft/20260519_023851
116
  - domain=general_knowledge
117
- - data=general_knowledge_rlvr
118
- - training=rlvr
119
- - training.temperature=1.0
120
- - training.num_generations=8
121
- - training.gradient_accumulation_steps=8
122
- - training.max_completion_length=256
123
- - training.learning_rate=1e-6
124
  - training.deepspeed=null
 
125
  job:
126
- name: rlvr
127
  chdir: false
128
- override_dirname: data=general_knowledge_rlvr,domain=general_knowledge,rlvr.base_model_dir=/scratch/checkpoints/general_knowledge/sft/20260519_023851,training.deepspeed=null,training.gradient_accumulation_steps=8,training.learning_rate=1e-6,training.max_completion_length=256,training.num_generations=8,training.temperature=1.0,training=rlvr
129
  id: ???
130
  num: ???
131
  config_name: base
@@ -139,22 +133,22 @@ hydra:
139
  runtime:
140
  version: 1.3.2
141
  version_base: '1.3'
142
- cwd: /scratch/runs/cs552-ammar-g07-rlvr-pure-210609
143
  config_sources:
144
  - path: hydra.conf
145
  schema: pkg
146
  provider: hydra
147
- - path: /scratch/runs/cs552-ammar-g07-rlvr-pure-210609/configs
148
  schema: file
149
  provider: main
150
  - path: ''
151
  schema: structured
152
  provider: schema
153
- output_dir: /scratch/checkpoints/general_knowledge/rlvr/20260520_190659/hydra
154
  choices:
155
  domain: general_knowledge
156
- training: rlvr
157
- data: general_knowledge_rlvr
158
  model: qwen3_17b
159
  hydra/env: default
160
  hydra/callbacks: null
 
112
  hydra:
113
  - hydra.mode=RUN
114
  task:
 
115
  - domain=general_knowledge
116
+ - data=general_knowledge_bare_mc_sft
 
 
 
 
 
 
117
  - training.deepspeed=null
118
+ - training.epochs=1
119
  job:
120
+ name: sft
121
  chdir: false
122
+ override_dirname: data=general_knowledge_bare_mc_sft,domain=general_knowledge,training.deepspeed=null,training.epochs=1
123
  id: ???
124
  num: ???
125
  config_name: base
 
133
  runtime:
134
  version: 1.3.2
135
  version_base: '1.3'
136
+ cwd: /scratch/runs/cs552-ammar-g07-sft-bare-043837
137
  config_sources:
138
  - path: hydra.conf
139
  schema: pkg
140
  provider: hydra
141
+ - path: /scratch/runs/cs552-ammar-g07-sft-bare-043837/configs
142
  schema: file
143
  provider: main
144
  - path: ''
145
  schema: structured
146
  provider: schema
147
+ output_dir: /scratch/checkpoints/general_knowledge/sft/20260519_023851/hydra
148
  choices:
149
  domain: general_knowledge
150
+ training: sft
151
+ data: general_knowledge_bare_mc_sft
152
  model: qwen3_17b
153
  hydra/env: default
154
  hydra/callbacks: null
hydra/.hydra/overrides.yaml CHANGED
@@ -1,10 +1,4 @@
1
- - rlvr.base_model_dir=/scratch/checkpoints/general_knowledge/sft/20260519_023851
2
  - domain=general_knowledge
3
- - data=general_knowledge_rlvr
4
- - training=rlvr
5
- - training.temperature=1.0
6
- - training.num_generations=8
7
- - training.gradient_accumulation_steps=8
8
- - training.max_completion_length=256
9
- - training.learning_rate=1e-6
10
  - training.deepspeed=null
 
 
 
1
  - domain=general_knowledge
2
+ - data=general_knowledge_bare_mc_sft
 
 
 
 
 
 
3
  - training.deepspeed=null
4
+ - training.epochs=1
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2ed1b2983ccab9627fe0cba395dbe44283bec458534ad72b428e25ab5dc2e439
3
  size 3441185608
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:41251f54c85a6b5eac3e543941bad3a5d7f1d56fb4c8763039fbb50258fb3eeb
3
  size 3441185608
tokenizer_config.json CHANGED
@@ -20,7 +20,7 @@
20
  "<|image_pad|>",
21
  "<|video_pad|>"
22
  ],
23
- "is_local": true,
24
  "local_files_only": false,
25
  "model_max_length": 131072,
26
  "pad_token": "<|endoftext|>",
 
20
  "<|image_pad|>",
21
  "<|video_pad|>"
22
  ],
23
+ "is_local": false,
24
  "local_files_only": false,
25
  "model_max_length": 131072,
26
  "pad_token": "<|endoftext|>",
training_args.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4741db8d2cb51430a6c04f08ac2a416b6a2f2d3ec90dec2047f38851a71b6d94
3
- size 7249
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:34815d74f6662a90112fe9becd3fda7c5403640e96e70671c8171023a080f476
3
+ size 5777