Levi980623

Qwen3-8B-E2C-SFT-only

README

License: apache-2.0

Recipe (identical to the 4B; see `code_changes/` and `training_records/report.md`)

Data: official 95/5 split (seed 42) → drop samples whose Execution is truncated at the source (supervised tokens ≥ 1995) → 49,684 train / 2,611 val; <|im_end|> appended to the target (solution_eos). The two training parquets are byte-identical to the 4B's (sha256 1f02662a… / 14a403f8…).
Trainer: E²C's verl fsdp_sft_trainer.py + the 4B patch (LR horizon = total_training_steps, periodic val loss, right-padding trim, no duplicate model build); verl/verl/models restored from verl v0.5.0 (8fdc4d3f).
Hyper-parameters: full-parameter FSDP2, bf16, gradient checkpointing, flash-attention 2; max_length 4096, truncation error; AdamW lr 2e-5, betas (0.9, 0.95), wd 0.01, warmup 10 %, cosine over 1000 steps, clip 1.0; global batch 64 (micro 1 per GPU, 4 × H200 → 16 accumulation); 1000 steps = 64,000 samples ≈ 1.29 epoch; seed 114514; save + val every 100 steps.
Environment: torch 2.8.0+cu128, transformers 4.56.2, flash-attn 2.8.3, tensordict 0.9.1, numpy 1.26.4, pyarrow 20, datasets 3.6 (Python 3.12). Wall-clock 2 h 04 min on 4 × H200 (6.8 s/step, 24.4 GB/GPU).

Results (same protocols as the 4B; details in `training_records/`)

Table with columns: 8B (this repo), 4B twin
	8B (this repo)	4B twin
step-1 train loss	0.467	0.521
full val loss step 800 / 900 / 1000 (2,611 records)	0.15884 / 0.15893 / 0.15893	0.18227 / 0.18235 / 0.18234
20 held-out MATH-500 (greedy): format / EOS stop / correct — step 800	20/20 · 20/20 · 19/20	—
same — step 1000	20/20 · 20/20 · 18/20	20/20 · 20/20 · 15/20
execution CE on non-restated tokens with / without the explicit plan (step 800)	0.140 / 0.210	≈ 0.158 / 0.230

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLMrepo = "Levi980623/Qwen3-8B-E2C-SFT-only"          # revision="step-1000" for the final-step weightstok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")prompt = tok.apply_chat_template([{"role": "user", "content": QUESTION}], tokenize=False, add_generation_prompt=True)ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)out = model.generate(**ids, max_new_tokens=4096, do_sample=False)   # stops at <|im_end|> (EOS was trained)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))

Chat template exactly as in training (no enable_thinking kwarg, no <think>). Unlike the official E²C recipe the model emits <|im_end|> after </EXECUTION>, so no stop string is needed.

Known issues shared with the 4B

bf16 parameters without fp32 master weights (official config): 51 RMSNorm gain tensors never changed.
>< in </EXPLORATION><EXECUTION> is a single token, so there is no clean token boundary between plan and execution.
Not the official exposure (≈1.29 epoch instead of ≈2.9); max_length, EOS and the source-truncation filter differ from the official script.
352 kept training samples have no \boxed{} (the filter is length-based only).

Files

model-0000x-of-00004.safetensors (bf16, untied lm_head), tokenizer, chat_template.jinja, generation_config.json; code_changes/ (launcher, trainer patch, tools, resolved config, Slurm scripts, requirements lock); training_records/ (report.md, metrics.json, loss curves, weight_diff_all.json, eval/ and eval_step800/ samples, val_full/ incl. empty-plan runs, preflight/ length statistics and loss-mask audit).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Levi980623

Model Tree

Base

Qwen/Qwen3-8B

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Recipe (identical to the 4B; see `code_changes/` and `training_records/report.md`)

Data: official 95/5 split (seed 42) → drop samples whose Execution is truncated at the source (supervised tokens ≥ 1995) → 49,684 train / 2,611 val; <|im_end|> appended to the target (solution_eos). The two training parquets are byte-identical to the 4B's (sha256 1f02662a… / 14a403f8…).
Trainer: E²C's verl fsdp_sft_trainer.py + the 4B patch (LR horizon = total_training_steps, periodic val loss, right-padding trim, no duplicate model build); verl/verl/models restored from verl v0.5.0 (8fdc4d3f).
Hyper-parameters: full-parameter FSDP2, bf16, gradient checkpointing, flash-attention 2; max_length 4096, truncation error; AdamW lr 2e-5, betas (0.9, 0.95), wd 0.01, warmup 10 %, cosine over 1000 steps, clip 1.0; global batch 64 (micro 1 per GPU, 4 × H200 → 16 accumulation); 1000 steps = 64,000 samples ≈ 1.29 epoch; seed 114514; save + val every 100 steps.
Environment: torch 2.8.0+cu128, transformers 4.56.2, flash-attn 2.8.3, tensordict 0.9.1, numpy 1.26.4, pyarrow 20, datasets 3.6 (Python 3.12). Wall-clock 2 h 04 min on 4 × H200 (6.8 s/step, 24.4 GB/GPU).

Results (same protocols as the 4B; details in `training_records/`)

Table with columns: 8B (this repo), 4B twin
	8B (this repo)	4B twin
step-1 train loss	0.467	0.521
full val loss step 800 / 900 / 1000 (2,611 records)	0.15884 / 0.15893 / 0.15893	0.18227 / 0.18235 / 0.18234
20 held-out MATH-500 (greedy): format / EOS stop / correct — step 800	20/20 · 20/20 · 19/20	—
same — step 1000	20/20 · 20/20 · 18/20	20/20 · 20/20 · 15/20
execution CE on non-restated tokens with / without the explicit plan (step 800)	0.140 / 0.210	≈ 0.158 / 0.230

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLMrepo = "Levi980623/Qwen3-8B-E2C-SFT-only"          # revision="step-1000" for the final-step weightstok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")prompt = tok.apply_chat_template([{"role": "user", "content": QUESTION}], tokenize=False, add_generation_prompt=True)ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)out = model.generate(**ids, max_new_tokens=4096, do_sample=False)   # stops at <|im_end|> (EOS was trained)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))

Chat template exactly as in training (no enable_thinking kwarg, no <think>). Unlike the official E²C recipe the model emits <|im_end|> after </EXECUTION>, so no stop string is needed.

Known issues shared with the 4B

bf16 parameters without fp32 master weights (official config): 51 RMSNorm gain tensors never changed.
>< in </EXPLORATION><EXECUTION> is a single token, so there is no clean token boundary between plan and execution.
Not the official exposure (≈1.29 epoch instead of ≈2.9); max_length, EOS and the source-truncation filter differ from the official script.
352 kept training samples have no \boxed{} (the filter is length-based only).

Qwen3-8B-E2C-SFT-only

README

Recipe (identical to the 4B; see code_changes/ and training_records/report.md)

Results (same protocols as the 4B; details in training_records/)

Usage

Known issues shared with the 4B

Files

Explore FriendliAI today

README

Recipe (identical to the 4B; see code_changes/ and training_records/report.md)

Results (same protocols as the 4B; details in training_records/)

Usage

Known issues shared with the 4B

Files

Recipe (identical to the 4B; see `code_changes/` and `training_records/report.md`)

Results (same protocols as the 4B; details in `training_records/`)

Recipe (identical to the 4B; see `code_changes/` and `training_records/report.md`)

Results (same protocols as the 4B; details in `training_records/`)