Full-parameter supervised fine-tune of Gemma-4-31B-it on multi-turn tool-calling / agent orchestration data, trained with plain cross-entropy and no loss shaping.

This model exists as the paired control for gemma4-31b-rovochat-full-sft-w1-260729, which used a shaped objective on identical data with an identical recipe. Released together so the contribution of the loss shaping can be measured rather than assumed.

⚠️ Neither this model nor its paired arm has been evaluated

These are raw training outputs. No benchmark, quality gate, or safety evaluation has been run on either. Training loss converged cleanly — that is not evidence of quality. In this model family, full-parameter fine-tuning is known to degrade open-ended text quality while tool metrics improve. Treat text quality as unverified and plausibly regressed versus the base model.

The controlled comparison

	W0 (this model)	W1 (treatment)
CE weight	1.0	1.0
first-decision term	0.0	4.0
terminal-contract term	0.0	4.0
KL anchor	0.0	0.0 (forced off under full-FT — not a difference)
CE normalization	`token_mean` (per supervised token)	`per_turn` (per turn)
final loss	0.9159	1.0160
loss minimum	0.6667 @ step 199	0.6900 @ step 199
wall-clock	44 h 25 m 51 s	44 h 25 m 41 s

Both arms bottom out at exactly step 199 — same seed, same data order, genuinely paired.

⚠️ The loss values above are NOT comparable across arms. token_mean averages over supervised tokens; per_turn averages over turns. They are different scales, so W0's lower number is an artifact of normalization, not evidence of quality. Only a downstream evaluation can compare them.

⚠️ This is a two-factor contrast. W0 changes the shaped weights and the CE normalization simultaneously, so a measured difference cannot be attributed to the shaping alone. W1's own decomposition puts its shaped terms at only ~5.6 % of the objective at convergence, so any gap is plausibly dominated by the normalization change.

Training


method	full-parameter FT (all non-vision params), DeepSpeed ZeRO-3, no offload, bf16
trainable	30.697 B / 31.273 B (98.16 %) — vision tower frozen
hardware	16 × H200 141 GB, 2 nodes, RoCE/RDMA
epochs	1 — 457 optimizer steps, global batch 16 (micro-batch 1 × 16 ranks, no accumulation)
optimizer	AdamW, LR 2e-6 cosine, warmup 3 %, weight decay 0.1, grad-clip 1.0
sequence	up to 44 k tokens, assistant-turn-masked loss
attention	SDPA (memory-efficient) — no FlashAttention
seed	17
objective	`loss = CE` exactly — verified: `total − ce` is `0.000e+00` on all 457 logged steps

Identical to the W1 arm in trainer, launcher, DeepSpeed config, dataset (all sha256-matched), hardware, and every hyperparameter. A field-by-field diff of the two runs' resolved configs differs only in the loss-config keys, the output directory, and a checkpoint-retention limit that has no effect on weights.

Data

Internal multi-turn tool-calling orchestration conversations (not released): 6,629 conversations, windowed to ≤44 k tokens → 7,310 training rows (3,441 whole; 3,869 window-pieces — split, never truncated) totalling 274.6 M tokens, of which 20.95 M (7.6 %) are supervised across 41,466 assistant turns. Loss is masked to assistant turns only.

Usage

python

from transformers import AutoModelForCausalLM, AutoProcessor
m = AutoModelForCausalLM.from_pretrained(
    "tzchen07/gemma4-31b-rovochat-full-sft-w0-260801",
    torch_dtype="bfloat16", device_map="auto", attn_implementation="sdpa")
p = AutoProcessor.from_pretrained("tzchen07/gemma4-31b-rovochat-full-sft-w0-260801")

Multimodal (VL) architecture with a frozen vision tower — all 1,188 tensors included (832 language, 355 vision, 1 projection). lm_head is absent by design (tied embeddings). processor_config.json is included: Trainer.save_model() does not write it, and the processor cannot be built without it.

Resuming

…-checkpoint-457 holds the complete end-of-training state: the same 2 safetensors shards, 16 ZeRO-3 fp32 optimizer shards, 16 RNG states, and trainer_state.json.

Note the optimizer state is only usable to resume this run at the same world size (16) and the same DeepSpeed configuration — it is not portable to other topologies. Resuming additionally requires the training script and dataset, which are not part of this release.

Limitations

Unevaluated (see banner). No capability, regression, or safety testing.
Trained on internal proprietary conversations; behaviour reflects one product's tool schema and will not transfer cleanly to other tool sets.
One epoch, single data source, no held-out validation loss was tracked during training.
Inherits all limitations and licence terms of the Gemma-4 base model.