GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
tomzhu0225
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
tomzhu0225
Model Tree
internlm/Intern-Decision-4B
Input Modalities
Output Modalities
Supported Functionality
This PEFT LoRA adapter specializes internlm/Intern-Decision-4B, pinned to
38c9dafbbef52dd04bae943c0c6ac918419adfd5, for proposing attention in autonomous
scientific workflows. The adapter weights have SHA-256
76bf1bdf4c6d6a261211057d3cab4c8609131788827569f9c33f802d3491dab5.
Its original adapter configuration has SHA-256
5ab93c62a751ccfe082c88caabd6c40c064d04442e3b1e164387690af6e8ba6a.
The pretrained base is required separately; this is not a standalone model.
The tested role is an assistant to a capable worker, director and independent repair judge. It proposes whether visible evidence warrants observing, checking again, requesting help, or escalating for review. It does not approve science, execute tools, grant repair authority, or replace the main scientific worker.
The scientific tests concern bounded asynchronous one-dimensional WarpX control and diagnostic/reporting workflows. They do not establish autonomous discovery, performance across scientific disciplines, or a production reliability claim.
The tested interface is the Simjecture scientific-attention 0.2 draft. Only
public state and decision questions enter the masked-chat input. Training
targets, future outcomes and hidden fixture labels remain outside that input.
The five scored fields are route, question_ref, basis, evidence_primary
and evidence_secondary. Evidence and question aliases are bound to canonical
identities by the host, rather than generated as arbitrary identifiers.
Routes combine attention with optional knowledge lookup or literature review. Such support requires a question supplied by the host. The plasma workflow tests declare no retrieval questions and therefore do not validate retrieval or literature-review utility.
Serving uses a single masked-position forward pass, temperature 1 and the original independent field decoder. The full-input allowance is 8,192 tokens; the server explicitly reports inputs beyond that allowance. Optional constrained route/question decoding is a separate implementation and is not part of this checkpoint's reported workflow comparison.
The tested workflow polls every five seconds while the main worker is parked, allows at most one automatic expert request, and retains a 90-second fallback. It does not establish the benefit of polling an actively reasoning worker.
The host must validate combinations and evidence, check freshness and current state, respect announced quiet periods, suppress duplicate work, and separately record actual dispatch. A proposal is not an admitted alert. Model option probabilities have not been calibrated as probabilities of scientific failure.
The initial supervised mixture contains 4,096 authored scientific examples,
400 procedural examples and 1,024 general decision examples. The selected run
performed 512 optimizer steps with accumulation 4 and learning rate 5e-5;
2,048 examples were processed, rather than a complete pass over that mixture.
LoRA rank is 16, alpha 32, with 32,464,896 trainable parameters.
This candidate then received eight continued-SFT updates with learning rate
1e-5 and accumulation 4. It is the continued-SFT control from the earlier
SFT/RL experiment, not the RL checkpoint. The later live-state reviewer
judgments and repeated-workflow outcomes have not been used to update it.
The two training stages total 2,080 example presentations; this is not a claim
that all 5,520 mixture rows were processed or that every presentation was unique.
The 1,024 generic rehearsal rows were constructed independently of the public
retention benchmarks. Actual training inputs ranged from 339 to 1,459 tokens;
the recorded live-state evaluation used 3,200–4,964-token inputs.
The full training mixture and all scientific evaluation splits were rebuilt from the pinned source and match the original byte hashes. An audit found no identical compiled visible inputs or shared IDs between training, development and test. The training mixture has 5,506 distinct visible inputs among 5,520 rows. Its constructed mechanisms remain shared across splits; this is not task-disjoint generalization. See the reproduction record for source, data and weight identities and the original training commands.
| Measure | Pretrained base | Continued SFT |
|---|---|---|
| Authored scientific contract/routing test | 54.0% | 97.5% |
| General-decision macro accuracy across seven suites | 90.0133% | 88.3426% |
| Route acceptable to both live-state reviewers | 16/24 | 18/24 |
| Accepted expert route on seven states where both require an expert | 0/7 | 5/7 |
| Expert proposal on eleven unanimous observe-only states | 0/11 | 2/11 |
The authored test contains 1,024 cases. General evaluation contains 10,751 input records and 12,351 scored decisions. The macro score weights the seven suites equally. Continued SFT passes the predeclared point-estimate tolerance of two percentage points of macro loss and three points per suite with at least 300 decisions. This does not show universal retention: it loses 8.11 points on the 111-example hard suite and 2.25 points on typed decisions. It should not be presented as a general-decision replacement with no regressions.
| Decision-formatted public suite | Input records | Scored decisions | Base accuracy | Continued SFT |
|---|---|---|---|---|
| JevBench easy | 48 | 48 | 100.00% | 100.00% |
| JevBench original | 72 | 72 | 98.61% | 98.61% |
| JevBench hard | 111 | 111 | 73.87% | 65.77% |
| ToolACE | 310 | 310 | 96.45% | 96.45% |
| Typed decisions | 400 | 2,000 | 80.45% | 78.20% |
| AG News | 7,600 | 7,600 | 90.84% | 90.91% |
| WildJailbreak | 2,210 | 2,210 | 89.86% | 88.46% |
These are option-scoring evaluations using the pinned JevBench scorer, not free-form agent executions. All reported static evaluations completed without inference errors. The exact retained metrics include dataset hashes and counts; this table is a documentation update, with no new model execution or optimizer update.
The live-state sample consists of 24 correlated observations from eight earlier episodes, judged by separate contexts of the same large model. These judgments are not human ground truth or causal intervention outcomes. The base observed throughout. SFT detects several problems but also proposes attention that both reviewers consider unnecessary. It routes two protected harness defects to the director where both reviewers expected the independent repair judge.
An earlier exploratory seven-episode-per-arm cohort measured means of 223.9 seconds for this adapter, 265.1 for rules and 277.2 for worker-only execution. It omitted the pretrained-base arm. These promising observed differences do not answer the missing base/SFT workflow comparison. The subsequently registered 64-workflow replication retained its initial verified rules episode and provider-failed base episode, then resumed only previously unstarted assignments. All 64 original assignments now have audit records: 58 verified completions, four infrastructure failures and two SFT workflow completion failures. The registered SFT-minus-base restricted-completion estimate is +12.6 seconds (95% interval −116.1 to +148.3); excluding the predefined whole infrastructure-affected blocks gives +11.9 seconds (−91.1 to +110.3). Neither establishes an overall advantage. SFT verifies 13/16 workflows; base, rules and worker-only execution each verify 15/16. Mean restricted completion times are 364.4, 351.8, 338.5 and 356.0 seconds, respectively. An unsuccessful workflow contributes 650 seconds, so these are not raw runtime means. Publication-retry and unstable-update workflows show favorable repeated SFT differences, while protected source selection fails in both SFT repeats. See the completed record for all arms, intervals, infrastructure sensitivity and individual paired effects.
Both R10 SFT source-binding workflows failed completion verification despite successful physical controls and authorized repairs. Their four main-worker turns were exhausted during replanning/help/reprocessing, leaving no final reports and one required return unacknowledged in each. These are retained as unsuccessful workflows under the registered limits, each receiving the 650-second restricted completion penalty. Model attention can consume coordination capacity as well as save it; these data do not establish that every early intervention helps a capable worker.
Inspection of recorded worker prompts found general research execution/review guidance alongside the experiment's different wrapper/report protocol. This common instruction context may influence discretionary help requests; its causal effect is unmeasured. A separate task-focused pilot without a decision model also reproduced a completion failure after routine help: corrected processing finished after the fourth worker turn, but the loop offered no fifth reporting turn despite unused wall time and native-start allowance. A prospective control permits additional resumptions within the same resource limits and is being commissioned. See the control record. These observations do not replace the original results or establish model benefit under a revised harness. Reported model measurements remain conditional on the original harness.
The tested runtime is a remote NVIDIA A100-SXM4-80GB, PyTorch 2.11.0 with CUDA
13.0, Transformers 5.20.1, PEFT 0.21.2, huggingface-hub 1.33.0 and FLA 0.6.0.
Use the pinned Simjecture input compiler and fla-prefill execution profile.
CPU, P40 and quantized inference performance are not validated for this checkpoint.
Reload verification checks all 496 adapter tensors, the disabled-adapter base
prediction and saved probability probes. Both SFT probes reproduced before and
after the recorded-state evaluation with maximum discrepancy below 2.85e-8
against the unchanged 1e-4 tolerance. The original weights remain preserved
off the disposable A100 runtime.
The pinned base README declares Apache-2.0. Its LICENSE and LICENSE-QWEN
files contain Apache License 2.0 with Shanghai AI Laboratory and Alibaba Cloud
copyright notices, respectively. Their exact text and hashes are retained in
the license evidence. The complete pinned
upstream listing contains no additional NOTICE or attribution file. Training-data
provenance and exact split reproduction are recorded separately. These checks
alone do not complete release.
A portable candidate copy preserves the weight-file hash and changes only the
base repository reference and pinned revision in its configuration. After timed
collection stopped for transport recovery, a separate fresh A100 load matched
all 496 adapter tensors exactly and reproduced all three saved base/SFT probes,
with maximum probability discrepancy 6.19e-8 against the unchanged 1e-4
tolerance. The reload evidence records this check.
The completed scientific comparison used the original adapter and configuration;
no optimizer ran during export or reload.
This release preserves the evaluated weight tensors and the pinned base reference.
The completed evidence supports an experimental research artifact, with the
limitations above. It does not establish overall workflow efficacy or production
reliability. Use the v0.1 tag or an immutable repository commit to reproduce this
checkpoint; later releases may contain different weights and evaluation results.
The training/retention report, live-state results, workflow protocol and retained evidence files are the current sources. This experimental release makes no claim of an overall scientific-workflow advantage.
