You can deploy this model and start earning money today!

pantomiman/
qwen3.5-0.8b-grpo-musique-m5_1_s43

Name: qwen3.5-0.8b-grpo-musique-m5_1_s43
Availability: InStock

safetensors

retrieval-augmented-generation

Views

No views yet

Qwen3.5-0.8B GRPO on MuSiQue: run m5_1_s43 (F1-only, seed 43)

One run of the M5 reward-shape x seed ablation (M9.5a). Qwen3.5-0.8B trained with GRPO + a Wikipedia search tool on MuSiQue; this run uses the F1-only reward at seed 43 on Blackwell.

field	value
run code	`m5_1_s43`
reward shape	F1-only
seed	43
hardware	Blackwell (DP=4 colocated vLLM)
checkpoints	23 (step_10..step_230, every 10 steps)
rollouts	230
best deployment ckpt (held-out avg EM)	step_210 (0.291 @step_210)
milestone	M9.5a

Reward: reward = f1(answer, gold); no floor, no format gate.

Repo structure (uniform across the 9 M5 runs)

step_<N>/policy/weights/model/consolidated/model-*.safetensors   # deployable consolidated model
step_<N>/policy/tokenizer/                                        # tokenizer (or backfill from Qwen/Qwen3.5-0.8B)
rollouts/train_data_step<N>.jsonl                                 # per-step training rollouts
config_snapshot.yaml                                             # GRPO config for this run

Family

Index + full 9-run matrix: pantomiman/qwen3.5-0.8b-grpo-musique-m5
Held-out 7-benchmark eval (uniform): pantomiman/reason-over-search-eval-m5 (m5_1_s43/...)
Collection: see the index repo. Recipe varies only by reward shape (m5_1=F1-only, m5_5=F1+format, m5_6=EM-only) x seed (42/43/44).

pantomiman/qwen3.5-0.8b-grpo-musique-m5_1_s43

Qwen3.5-0.8B GRPO on MuSiQue: run m5_1_s43 (F1-only, seed 43)

Repo structure (uniform across the 9 M5 runs)

Family

pantomiman/
qwen3.5-0.8b-grpo-musique-m5_1_s43